Theory of conciousness DPSH
Dynamic Perceptual State Hypothesis
DPSH
Michal Havelec
version 1.0.0
2026-08-20
Contents
-
- Introduction to DPSH
-
- Formal model of a dynamic neuronal system
-
- Spontaneous stochastic activity
-
- Endogenous oscillators as a temporal structure
-
- Non-commutative neuronal dynamics
-
- Self-organization and spontaneous symmetry breaking
-
- The metastable Perceptual Manifold
-
- Hysteresis and the continuity of the inner world
-
- Predictive constraint on the dynamics
-
- Learning the dynamics and Deep State Learning
-
- Global Workspace and the global availability of the percept
-
- The Perceptual Manifold as a shared internal model of the world
-
- Phenomenal state and qualia
-
- Integrated predictions and the falsification program for the entire DPSH
-
- DPSH requirements for the Cognia engine and the experimental framework
1. Introduction to DPSH
Dynamic Perceptual State Hypothesis
DPSH
1.1 The problem of percept formation
Contemporary neural models can classify sensory inputs with high accuracy, build representations of them, predict future inputs and generate corresponding responses. The mere ability to transform input information into output does not, however, explain how a continuous internal state corresponding to the currently perceived world can arise in a neuronal system.
It is therefore necessary to distinguish between several levels of information processing:
- detection – the system responds to a particular property of the input,
- representation – the system creates an internal state correlating with a particular property or object,
- percept – the representation becomes part of a continuous integrated state of the system and influences its further dynamics,
- phenomenal experience – the subjective aspect of the percept, that is, what is often referred to in philosophy of mind as qualia.
This hypothesis is primarily concerned with the third level: the formation and maintenance of the percept.
Phenomenal experience constitutes a stronger problem. The hypothesis therefore does not assume that finding a mechanism of the percept automatically explains the existence of qualia. It does, however, examine the possibility that a dynamic neuronal mechanism producing a percept may constitute a candidate physical substrate on which phenomenal experience depends.
1.2 From representation to dynamic state
A conventional neural network can be understood, in simplified terms, as a transformation
X -> F(X) -> Y
where X represents the input, F the neuronal computation and Y the
resulting representation or output.
For a living perceptual system, however, such a description is not sufficient. An organism does not begin the processing of each new sensory input from a zero state. At every moment a certain internal state already exists, one that arose from the system's previous interaction with the environment.
A more appropriate description is therefore
S(t + dt) = F(S(t), I(t))
where:
S(t)is the current internal state of the system,I(t)is the sensory and internal input,Frepresents the dynamics of the neuronal system.
The same sensory input therefore need not always produce the same resulting state:
F(S_A, I) != F(S_B, I)
On this account, perception is a process dependent on the history of the system.
The system does not create isolated representations of individual moments. It continuously transforms an already existing internal model.
1.3 The percept as a dynamic process
The Dynamic Perceptual State Hypothesis proceeds from the assumption that a percept is not necessarily represented by the activation of a particular neuron, of a neuronal population, or by a static pattern of activity.
Consider the global state of a neuronal system
S(t) = (s1(t), s2(t), ..., sn(t))
where si(t) represents the state of individual neuronal or other
dynamic units.
A percept may correspond to a region M of the state space in which the
trajectory of the system moves for a certain period of time:
S(t) in M
The individual components of this state may change continuously:
si(t) != si(t + dt)
while the global dynamic structure is preserved.
A percept therefore need not be a static state.
It may be a metastable dynamic process.
This difference is fundamental. Information need not be retained solely through unchanging activation. It may also be retained by the structure of the system's trajectory in its state space.
1.4 Continuous internal representation of the world
A perceptual system must solve a fundamental problem: sensory information is incomplete, delayed, burdened with noise and constantly changing.
Nevertheless, we do not subjectively perceive the world as a sequence of independent sensory samples.
We perceive a relatively stable environment.
An object, for example, does not cease to be part of our perceived world merely because it is briefly occluded by another object.
This suggests that the perceptual system does not work only with current sensory data, but maintains an internal state against which new data are interpreted.
This principle is compatible with predictive processing, according to which perception involves an ongoing interaction between an internal model and sensory evidence.
Schematically:
internal state
|
v
prediction
|
v
expected input
|
| comparison
v
sensory input
|
v
prediction error
|
v
modification of internal state
Within DPSH, however, predictive processing does not constitute a complete explanation of percept formation.
It primarily determines the constraints on the system's dynamics: sensory evidence favours certain possible internal states and destabilizes others.
The question remains by what physical and neuronal mechanism the dynamic internal state itself arises.
1.5 Global clocklessness
One of the main assumptions of DPSH is that a biological neuronal system cannot be fully characterized as a network updated by a shared global computational step.
Individual neurons are autonomous dynamic units.
They receive signals at different moments, change their internal state and generate further events without requiring that all other neurons simultaneously perform the same computational step.
DPSH therefore assumes globally clockless, locally causal dynamics.
This does not mean the absence of temporal organization.
A neuronal system may contain oscillations, and individual local circuits may be markedly synchronized.
It is necessary, however, to distinguish between:
global processing clock
and
endogenous oscillatory signal
A global clock determines when the system may be updated.
An endogenous oscillator, by contrast, is part of the system itself. Its signal may change the excitability of neurons, the probability of a spike, plasticity or communication between populations, but it does not determine a universal moment of update for all neurons.
The temporal structure of the computation therefore need not be imposed on the system from outside. It may arise within its own dynamics.
1.6 Stochasticity as part of the computation
The second assumption of the hypothesis is that the variability of neuronal activity need not represent merely error or unwanted noise.
A neuron may have a non-zero probability of a spike even in the absence of an unambiguous external stimulus:
P(spike | external_input = 0) > 0
Its instantaneous probability of activity can be understood in general as a function
P(spike_i, t) =
F(
sensory_input,
internal_state,
recurrent_input,
oscillatory_phase,
synaptic_history,
stochastic_component
)
Such a network remains dynamic even in the absence of an immediate sensory stimulus.
DPSH examines the possibility that limited stochasticity allows the system to explore nearby regions of its own state space.
Recurrent connections, local oscillations, plasticity and sensory evidence may then amplify, stabilize or destabilize some of these states.
Stochasticity on this account is not the opposite of structure.
It may be one of the mechanisms out of which structure self-organizes.
1.7 Time as a carrier of information
If neurons are not updated by a shared global clock, the relative timing of events becomes a potentially significant part of the computation.
Information is then not determined only by
which neurons spiked
or
how many times they spiked,
but also by
when they spiked
and with respect to what local dynamic context.
This leads to an important consequence.
Neuronal transformations may be non-commutative:
F_B(F_A(S)) != F_A(F_B(S))
The sequence of events
A -> B
therefore need not lead to the same internal state as
B -> A.
The history of the system thereby becomes a physical part of its present state.
Spike-timing-dependent plasticity is a well-known example of a mechanism in which the relative order of events influences the change of synaptic connections. DPSH examines the broader possibility that order-dependent dynamics is a fundamental property of the formation of perceptual states itself.
1.8 Self-organization of the percept
Sensory input need not unambiguously determine a single interpretation.
At a given moment the dynamics of the system may admit several competing states:
M1, M2, ..., Mn
Sensory evidence changes their stability, but need not itself contain an unambiguous decision as to which of them is to be realized.
DPSH assumes that through recurrence, stochasticity, excitation, inhibition and temporal coordination a spontaneous breaking of this dynamic symmetry may occur:
competing possible states
|
v
local fluctuations
|
v
recurrent amplification
|
v
symmetry breaking
|
v
metastable perceptual state
A coherent percept therefore need not be the result of a central mechanism that explicitly assembles individual pieces of sensory information.
It may arise as a macroscopic property of a self-organizing dynamic system.
1.9 Perceptual Manifold
For a working description of the global internal state we introduce the notion of the Perceptual Manifold.
The Perceptual Manifold is not meant as a particular anatomical region nor as a single neuronal representation.
It denotes a dynamic structure of the system's state space in which mutually dependent information about the current internal model of the environment and of the organism is simultaneously encoded.
Individual subsystems may obtain different information from this state:
Perceptual Manifold
|
+----+----+---------+----------+
| | | |
v v v v
action memory valuation language
|
v
environment
The perceptual state therefore need not be the final output of the perceptual system.
It is at the same time an input for further neuronal processes.
This gives rise to a closed dynamic loop:
environment
|
v
sensory input
|
v
perceptual dynamics
|
v
internal model
|
+----> prediction
|
+----> memory
|
+----> evaluation
|
+----> action
|
v
environment
The system thus continuously changes the world from which it subsequently obtains further sensory data.
1.10 Relation to Global Workspace Theory
DPSH does not understand the Global Workspace as a mechanism that necessarily creates the percept itself.
It proposes to distinguish between:
percept formation
and
global accessibility.
The working architecture is:
local sensory dynamics
|
v
metastable perceptual state
|
v
workspace selection
|
v
global broadcast
|
+-------+-------+
| | |
v v v
memory action cognition
The Global Workspace may explain how a certain content becomes globally available to other processes.
DPSH attempts to address the preceding question:
How can a coherent dynamic content that is to be made globally accessible arise in the first place?
1.11 The central hypothesis
On the basis of the preceding assumptions we formulate a working central hypothesis:
In a globally clockless recurrent neuronal system, autonomous stochastic spiking units may, through local interactions, endogenous oscillatory modulation, relative timing, synaptic delays and local plasticity, spontaneously create metastable population states. Sensory evidence and predictive mechanisms constrain the dynamics of these states such that some of them create persistent internal representations of relevant properties of the environment. These dynamic states constitute a candidate mechanism for the formation of an integrated percept.
1.12 The strong phenomenal hypothesis
Above this mechanistic hypothesis a stronger hypothesis can be formulated:
A metastable dynamic perceptual state may constitute the neuronal substrate of the phenomenal content of experience.
This hypothesis cannot, however, be derived merely from the existence of metastable neuronal dynamics.
Nor does a successful experimental demonstration of a dynamic perceptual state by itself prove the emergence of subjective experience or qualia.
For this reason, three epistemic levels of claim, designated E1–E3,
will be strictly separated in further research, so that they are not
confusable with the partial hypotheses H1–H13 of the individual
chapters:
E1: existence of a dynamic mechanism
E2: capacity of the mechanism to create and maintain perceptual content
E3: relation of this mechanism to phenomenal experience
E1 and E2 can be tested directly through neuronal simulations and behavioural tasks.
E3 remains an open phenomenal hypothesis.
1.13 Falsifiability
The aim of DPSH is not to create a mechanism that can be retrospectively adapted to any result whatsoever.
The individual parts of the hypothesis must generate measurable predictions.
If, for example:
- removing asynchronous processing does not change the relevant dynamics,
- removing spontaneous stochasticity does not change the capacity to form internal states,
- disrupting the relative phase while preserving the firing rate does not change the perceptual representation,
- the order of neuronal events does not have the expected influence on the state trajectory,
- the network does not exhibit metastable population structures,
- the previous state does not influence the interpretation of a subsequent input,
- or predictive feedback does not stabilize states corresponding to the structure of the environment,
then the corresponding parts of the hypothesis will be weakened or falsified.
The research goal is therefore not to create a system that merely exhibits interesting behaviour, but to determine experimentally which of the proposed mechanisms are genuinely causally necessary for the formation of a dynamic perceptual state.
2. Formal model of a dynamic neuronal system
2.1 Basic assumption
The Dynamic Perceptual State Hypothesis does not regard a neuronal network primarily as a sequence of discrete transformations
S(t) -> S(t + 1)
governed by a shared global step.
The basic object of the system is an autonomous dynamic unit that exists in time, maintains its own internal state, receives events from other units and may, depending on its state, generate new events.
The network as a whole therefore has no single moment at which it is completely updated.
Its global state is the emergent result of locally occurring processes running concurrently.
For a system composed of N units, the global state can be written
formally as
S(t) = {s1(t), s2(t), ..., sN(t)}
where si(t) represents the internal state of unit i at time t.
Time t here does not represent the number of a global computational
step.
It is merely a coordinate with respect to which causal relations between events can be described.
2.2 The autonomous neuron
A neuron Ni is a dynamic unit with its own state
si(t)
which may contain, for example:
membrane potential
refractory state
activation history
adaptation state
stochastic state
local modulatory state
Exact biological fidelity of the individual parameters is not the goal in the first phase of the research.
What matters are the properties of the system:
- the neuron has a state persisting in time,
- its state may be changed by an incoming event,
- its state may change even without an external event,
- it may generate a spike,
- the probability or the moment of a spike may depend on its history,
- its update does not require the simultaneous update of other neurons.
In general its dynamics can be described as
dsi/dt = Fi(
si(t),
Ii(t),
Ri(t),
Oi(t),
Hi(t),
ξi(t)
)
where:
Ii(t)is the external input,Ri(t)is the recurrent synaptic input,Oi(t)is the oscillatory/modulatory input,Hi(t)represents the relevant history,ξi(t)is the stochastic component.
A neuron is therefore not merely a function of the present input.
Its response depends on its current dynamic state.
2.3 The spike as an event
The basic unit of communication is the spike.
We understand a spike as a temporally localized event
ei = (i, t)
meaning that neuron i generated a spike at time t.
The spike itself need not carry a scalar value analogous to the activation of a conventional artificial neuron.
Information may be contained in:
the identity of the source,
the number of spikes,
the frequency of spikes,
the relative timing,
the order of spikes,
the phase with respect to a local oscillation,
the relation to the activity of other neurons.
The basic carrier of information is therefore not necessarily an isolated spike.
It may be the spatiotemporal structure of a set of events
E = {e1, e2, ..., en}.
2.4 Global clocklessness
The system does not contain a mechanism of the type
for every neuron:
calculate next state
followed by a global
commit next state.
Such a mechanism would create a discrete sequence of global states
S0 -> S1 -> S2 -> ... -> Sn.
DPSH instead assumes local causality.
If neuron A emits a spike at time
tA
and the signal needs a time
dAB,
to reach neuron B, the event may influence neuron B at time
tB = tA + dAB.
Neuron C may in the meantime be influenced by an entirely different
event.
There is no requirement that
tB = tC.
The global state of the system therefore changes as a consequence of temporally dispersed local events.
2.5 Local causality
Every change of state must have a local causal history.
If
si(t1) != si(t0),
it must be possible to determine the mechanism by which the change arose:
previous internal dynamics,
incoming spike,
stochastic transition,
oscillator/modulator,
external sensory event,
plasticity event.
This principle is important for experimental interpretation.
A global perceptual state must not be created by an engine operation of the type
create_global_state(...).
It must arise as a consequence of local interactions.
Otherwise the system would presuppose precisely the mechanism whose emergent formation the hypothesis is attempting to explain.
2.6 The synapse as a temporally oriented connection
A synapse between neurons i and j is not merely a scalar weight.
The minimal abstract representation is
Wij = {
weight,
delay,
plasticity_state
}
A spike of neuron i
ei(t)
therefore does not immediately produce a change in neuron j.
It creates a causal event
ei(t)
->
Wij
->
ej_input(t + dij)
where dij is the synaptic and/or axonal delay.
The delay is not regarded merely as an implementation imperfection.
In a globally clockless system it may become part of the computation.
Two paths
A -> B -> D
and
A -> C -> D
may have different total delays.
Their signals may therefore, in neuron D:
overlap in time,
miss each other,
interfere,
amplify,
or influence different phases of its dynamics.
The topology of the network thus simultaneously creates a temporal topology.
2.7 Stochasticity of the neuron
A neuron need not have a deterministic boundary
activation > threshold -> spike.
A more general model allows
P(spike_i(t)) =
F(si(t), Ii(t), Ri(t), Oi(t), ξi(t)).
Even with minimal input it may hold that
P(spike_i) > 0.
We do not automatically regard this baseline firing as an error.
It is an experimental question whether it allows the network to explore the state space and whether it thereby contributes to the formation of dynamic representations.
At the same time, the hypothesis does not assume that more stochasticity is always better.
We assume the possibility of a relation
too little stochasticity
-> rigid dynamics
intermediate stochasticity
-> exploration + structure
too much stochasticity
-> loss of coherence.
The existence of such an optimum must be verified experimentally.
2.8 Spontaneous activity and the absence of input
An important consequence of the preceding point is:
sensory input = 0
does not imply
neural dynamics = 0.
The network may exhibit spontaneous activity.
Its present state may influence the probability of future spontaneous events, so that:
S(t)
->
spontaneous activity
->
S(t + dt).
The network thereby continues in its own dynamics even without new sensory information.
This mechanism is a candidate for:
exploration of internal state space,
maintenance of metastable states,
spontaneous transitions,
consolidation of learned dynamics.
The last point is so far hypothetical and must be tested experimentally.
2.9 The endogenous oscillator
An oscillator is an independent dynamic element of the system.
It can be described abstractly, for example, as
Ok(t) = Ak * sin(ωk*t + φk)
or by means of another periodic or quasiperiodic dynamic mechanism.
More important than the particular mathematical form are its architectural properties.
The oscillator:
- exists inside the neuronal system,
- creates a locally available signal,
- may influence neuronal excitability,
- may modulate the probability of a spike,
- may influence plasticity,
- may be influenced by other parts of the system.
The oscillator does not, however, determine:
"now update all neurons."
Therefore it holds that
endogenous oscillator != global processing clock.
2.10 Multiple temporal references
The network may contain a set of oscillators
O = {O1, O2, ..., Om}
with different:
frequencies,
phases,
amplitudes,
spatial ranges,
coupling strengths.
A neuron may be influenced by several oscillators:
P(spike_i, t) =
F(
...,
O1(t),
O3(t),
O7(t)
).
The significance of a spike may then depend not only on the absolute moment of its occurrence, but on its relative position with respect to several local temporal structures.
For example:
phase(O1) = 0.2π
phase(O3) = 1.4π
may represent a different dynamic context than
phase(O1) = 1.2π
phase(O3) = 0.4π
even if the instantaneous number of spikes is the same.
This gives rise to the possibility of a temporal code that is not based on a global clock.
2.11 Resonance and selective communication
If the excitability of a neuron depends on the oscillatory phase, the same spike need not have the same effect at all moments.
It may approximately hold that
response(spike, phase_A)
!=
response(spike, phase_B).
Two neuronal populations may therefore communicate more efficiently in certain mutual phase configurations.
A communication channel is not necessarily firmly open or closed.
Its effective throughput may be a dynamic function of time:
Cij(t) = F(φi(t), φj(t), ...).
This allows the same anatomical network to create different functional networks at different moments.
2.12 Non-commutativity of the dynamics
Because the state of a neuron depends on history and events arrive at different times, it cannot in general be assumed that
A(B(S)) = B(A(S)).
On the contrary, we expect
A(B(S)) != B(A(S)).
The arrival of events
A -> B
may lead to a different state than
B -> A.
Non-commutativity may arise at least through:
membrane dynamics,
refractory periods,
synaptic integration,
adaptation,
phase dependence,
STDP,
recurrent feedback.
The consequence is that a neuronal system cannot be fully characterized merely by the set of events that occurred in it.
It is also necessary to know their causal and temporal ordering.
2.13 Time-dependent plasticity
A synaptic weight is not necessarily constant:
wij = const.
In general:
wij(t + dt) =
wij(t) + Δwij.
One of the mechanisms may be dependence on the relative timing of pre- and postsynaptic activity:
Δt = t_post - t_pre
and
Δwij = G(Δt, local_state, modulators, ...).
The history of spikes thereby acquires the capacity to change the very dynamic space in which future interactions will take place.
The network therefore simultaneously:
evolves within its state space
and
modifies the structure of its state space.
This is crucial for learning.
2.14 The memory cell
DPSH does not exclude explicit state or memory elements.
A memory unit may, for example, maintain a state
M(t) = c
until a certain event causes
M(t) -> c'.
Such units may be useful for working memory, control, sequential tasks or the explicit retention of information.
The hypothesis, however, distinguishes between:
stored state
and
emergent dynamic state.
A memory cell retains an explicit value.
A metastable population state may retain a structure without any individual unit having to hold its complete representation.
This distinction will be experimentally important.
The aim is not to prove that explicit memory elements are unnecessary.
The aim is to find out which kinds of internal representation can arise from distributed dynamics alone.
2.15 Local versus global state
No neuron need contain information about the complete global state
S(t).
A neuron has access only to a limited set of information:
local state,
incoming connections,
modulatory signals,
local oscillations,
internal history.
The global state
S(t)
is therefore an analytical description of the system by an observer, not necessarily an explicit data structure available to the neurons.
This is a fundamental requirement.
If the Cognia engine maintained an object
GlobalPercept
and the neurons read directly from it, we would not be creating an emergent percept.
We would merely be implementing it as a hidden central variable.
2.16 The formation of a macrostate
Let us assume a large number of local units
N1 ... NN.
Their interactions may create collective quantities that are not a property of any individual unit.
Let us denote such a quantity
Ψ(S).
Ψ may characterize, for example:
population coherence,
cluster membership,
phase organization,
attractor occupancy,
metastable state identity.
DPSH assumes that perceptually relevant information may exist precisely at this macroscopic level.
By analogy with the order parameter in a physical system, the global structure need not be explicitly represented by a single component.
It arises collectively.
2.17 Symmetry breaking
In some situations several global configurations may represent similarly probable dynamic possibilities:
M1 ~ M2 ~ M3.
Small local fluctuations may be amplified through recurrence:
fluctuation
->
local advantage
->
recurrent amplification
->
population reorganization.
The result may be
M1 >> M2, M3.
Such a process constitutes a candidate mechanism for the spontaneous selection of one of several possible perceptual interpretations.
The selection need not be carried out by any central controller.
2.18 The metastable state
The resulting state need not be a permanent attractor.
We define a metastable state M as a region of the state space in which
the dynamics of the system remains for a limited time, even though the
individual components continue to change.
S(t) in M
for
t0 < t < t1.
A transition may then occur
M_A -> M_B.
The important measurable properties will be:
lifetime,
internal variance,
transition probability,
separability,
robustness to perturbation,
dependence on sensory input,
dependence on previous state.
In DPSH the percept is associated, as a candidate, precisely with this level of dynamic organization.
2.19 Hysteresis
If the present state depends on the previous trajectory, we expect hysteresis.
Upon a change of input
A -> B
the transition need not occur at the same point as upon
B -> A.
Formally:
transition_threshold(A -> B)
!=
transition_threshold(B -> A).
Hysteresis provides an experimentally measurable indicator that the network does not create a representation merely as an instantaneous function of the input.
Its interpretation depends on history.
2.20 Predictive constraint on the dynamics
Spontaneous dynamics creates many possible states.
Not all of them correspond to the environment.
We therefore introduce a predictive mechanism.
The internal state generates an expectation
P(t + dt) = G(S(t)).
The sensory system subsequently provides
I(t + dt).
A locally realized deviation arises
ε = I - P.
DPSH does not assume that there must exist a single central scalar
global_prediction_error.
The prediction error may be distributed among many local circuits.
Its function is to modify the dynamics so that states persistently incompatible with sensory evidence lose stability.
Prediction therefore does not create the percept directly.
It acts as a constraint on the space of states that can survive in the long run.
2.21 The dynamic perceptual state
On the basis of the preceding definitions we can characterize a working
perceptual state as a metastable macrostate M that satisfies several
properties:
- it arises through distributed local dynamics,
- it is not explicitly stored in a single unit,
- it persists across changes of individual neuronal activities,
- it is influenced by sensory evidence,
- it depends on the previous state of the system,
- it may influence subsequent processing,
- it may be available to several functional subsystems,
- it can be distinguished from other states by means of population dynamics.
This definition does not yet contain a requirement of phenomenal experience.
It defines only a candidate mechanistic substrate of the percept.
2.22 Perceptual Manifold
The set of dynamic perceptual states and of the transitions between them creates a structured region of the global state space.
We designate this structure provisionally as the
Perceptual Manifold.
It can be understood as
M = {
perceptual states,
trajectories,
transition probabilities,
learned constraints
}.
The Perceptual Manifold is therefore not a static map of the world.
It is a dynamic structure of the possibilities along which the internal state of the system may develop.
Learning may change its geometry:
experience
->
plasticity
->
altered state-space geometry
->
altered future perception.
Familiar or repeatedly experienced structures of the environment may thus correspond to regions of the dynamics into which the system passes more easily or which are more stable.
2.23 Requirements for the Cognia engine
The formal model gives rise to the first set of implementation requirements.
Cognia must experimentally allow at least:
- autonomous stateful neurons,
- event-driven propagation of spikes,
- synaptic delays,
- stochastic baseline firing,
- parameterizable excitability,
- refractory period,
- local oscillators,
- modulation of a neuron by an oscillator,
- multiple oscillators with different phases and frequencies,
- local timing-dependent plasticity,
- recurrent connectivity,
- excitatory as well as inhibitory interactions,
- explicit memory cells as a separate type of element,
- continuous running without a reset between individual inputs,
- recording of exact spike times,
- recording of the internal states of neurons,
- recording of synaptic changes,
- the possibility of experimentally switching off individual mechanisms,
- the possibility of phase scrambling,
- the possibility of timing scrambling while preserving the spike count,
- the possibility of a synchronous control regime,
- export of the global state-space trajectory for subsequent analysis.
The last point is especially important.
The engine must not merely provide the resulting output.
The object of research is the trajectory itself:
S(t0) -> S(t1) -> ... -> S(tn).
Without measuring it, it will not be possible to decide whether the assumed metastable structures genuinely arise.
2.24 Basic experimental principle
It must be possible to compare every mechanism with a control variant.
For example:
stochastic ON vs stochastic OFF
oscillators ON vs oscillators OFF
phase intact vs phase scrambled
asynchronous vs synchronous
STDP ON vs STDP OFF
recurrence ON vs recurrence reduced
history intact vs state reset.
An important principle will be to control the other quantities.
If, for example, phase scrambling simultaneously reduces the firing rate dramatically, it cannot be concluded from the result that the cause of the change was the phase.
We will therefore endeavour to construct experiments of the type:
same input
same topology
same approximate firing rate
same approximate spike count
same network capacity
but
different temporal organization.
In this way the causal significance of individual dynamic properties can be tested.
2.25 Chapter research hypothesis
The formal model leads to a first general experimental question:
Can a system composed solely of locally interacting autonomous units without a global update clock spontaneously create reproducible metastable macrostates whose identity, stability and transitions carry information about the sensory history of the system?
We formulate this question as the partial hypothesis H1:
H1 – Globally Clockless Dynamics Hypothesis
A neuronal system formed by autonomous units that maintain their own internal state, communicate by discrete events and update without a shared global update step may, through local causality and temporally oriented connections, create reproducible global macrostates. The absence of a global clock is not merely an implementation detail, but the condition under which the temporal structure of events is a carrier of information.
If not, the basic mechanistic assumption of DPSH will have to be fundamentally revised.
If so, the following chapters must determine which mechanisms are genuinely necessary for the formation of these states and whether they can fulfil the function of a perceptual representation.
3. Spontaneous stochastic activity
3.1 An active network without external input
One of the basic assumptions of the Dynamic Perceptual State Hypothesis is that the absence of new sensory input does not mean the absence of neuronal dynamics.
Formally:
I_external(t) = 0
does not imply:
dS(t)/dt = 0.
A neuronal system may remain active even at a moment when no significant change of the external environment is taking place.
This property is important for the hypothesis of a continuous percept.
If internal dynamics existed only as an immediate reaction to external input, then it would be natural to describe perception as a sequence of transformations:
input
->
processing
->
output
->
inactivity.
DPSH assumes a different regime:
previous state
|
v
ongoing dynamics <---- sensory input
|
v
new state
|
+-------------> ongoing dynamics
Sensory input therefore does not create the dynamics of the system from the beginning.
It modifies dynamics that already exist.
3.2 Spontaneous activity and stochasticity are not the same
It is necessary to distinguish two related but different notions.
Spontaneous activity means that a neuron or a neuronal population may exhibit activity without an immediate external stimulus.
Stochasticity means that the evolution of the system is not fully deterministic under the same macroscopic conditions.
Spontaneous activity may be deterministic.
For example, an oscillator may generate regular activity:
spike -> wait -> spike -> wait -> spike
without any external input.
Conversely, a stochastic neuron may have a probability of a spike:
P(spike_i, t) = p_i(t)
and the particular moment of the spike is not unambiguously determined in advance.
DPSH assumes the presence of both properties:
spontaneous activity
+
stochasticity.
Their functions must, however, be distinguished experimentally.
3.3 Baseline firing
For a neuron Ni we define a basic probability of spontaneous activity
p0_i > 0.
This represents the baseline firing.
It does not mean that the neuron must emit a spike regularly.
It means only that even without significant external input there is a non-zero probability of an event:
P(spike_i | I_external = 0) > 0.
The current probability may be modified by the state of the neuron:
P(spike_i, t) =
F(
p0_i,
membrane_state_i,
recurrent_input_i,
oscillatory_input_i,
adaptation_i,
history_i
).
Baseline firing is therefore not an independent generator of random spikes.
It is one of the components of the neuron's dynamics.
3.4 The spike as a local perturbation of the system
A spontaneous spike can be understood as a small local perturbation.
The neuron:
Ni
generates an event:
ei(t).
This may propagate through synapses:
Ni
|
+----> Nj
|
+----> Nk
|
+----> Nl.
In most cases a spontaneous spike may die out quickly without a macroscopic consequence.
In a different dynamic context, however, the same event may reach a system that is close to a transition between two states:
M_A ~ M_B.
A small fluctuation may then be amplified by recurrence:
stochastic spike
->
local perturbation
->
recurrent amplification
->
population transition
->
M_B.
Stochasticity thus provides a mechanism by which the system may spontaneously pass between the available regions of its state space.
3.5 Stochasticity as exploration of the state space
Let us assume that the network has a set of potentially available states:
Ω = {S1, S2, ..., Sn}.
Purely deterministic dynamics may, under the same initial conditions, repeatedly follow the same trajectory:
S0 -> S1 -> S2 -> S3 -> ...
Slight stochasticity allows deviations:
-> S2a -> ...
/
S0 -> S1 -> S2
\
-> S2b -> ...
The system may thereby visit alternative regions of the state space.
DPSH designates this possible function as:
state-space exploration.
This does not mean that a neuron or a network actively "searches" for a solution.
Exploration is a macroscopic consequence of local stochasticity.
3.6 Stochasticity and the learning of internal states
This property may be significant in combination with plasticity.
Let us assume a spontaneous trajectory:
S_A -> S_B -> S_C.
If local plasticity occurs during this trajectory:
spike_i
+
spike_j
+
Δt
->
Δw_ij,
the internal dynamics itself may gradually change the future probability of transitions.
A feedback loop thereby arises:
spontaneous dynamics
|
v
plasticity
|
v
modified connectivity
|
v
modified dynamics
|
+----------------+
|
v
further plasticity.
This is one of the candidate mechanisms for what we provisionally designate as Deep State Learning.
The strong version of the hypothesis says:
A network need not learn only the relation between an external input and a required output. Through its own ongoing activity it may change the probability and the stability of its internal dynamic states.
This claim is not accepted as fact in DPSH.
It is an experimental hypothesis.
3.7 Reactivation of experience
If previous experience has changed the synaptic structure of the network, spontaneous activity no longer takes place in the original state space.
Learning has changed its dynamics:
Ω_before
->
experience
->
plasticity
->
Ω_after.
Spontaneous activity in Ω_after may therefore preferentially visit
trajectories similar to the states created by previous experience.
Schematically:
external experience
A -> B -> C
|
v
plasticity
|
v
spontaneous dynamics
A' -> B' -> C'.
This need not be an exact reproduction of the original spikes.
What may matter is the similarity of the macroscopic trajectory.
DPSH therefore assumes the possibility that spontaneous activity enables repeated internal reactivation of structures created by experience.
If plasticity remains active during this activity, such reactivation may further change the dynamics of the network.
3.8 Spontaneous activity as prevention of dynamically dead states
Under certain conditions a network without spontaneous activity may end up in a state:
S_dead
for which:
no input
->
no spike
->
no state transition
->
no plasticity.
Such a state is stable, but computationally uninteresting.
DPSH examines the possibility that slight baseline activity reduces the probability of becoming stuck in similar dynamically dead states.
A spontaneous spike may:
disturb S_dead
->
activate local circuit
->
expose new synaptic interaction
->
initiate new trajectory.
This property may be especially significant for a system whose learning depends on local activity.
Synapses that are never activated have no opportunity to participate in timing-dependent plasticity.
Spontaneous activity may therefore potentially provide a low level of ongoing "testing" of the existing connections.
3.9 Why maximal stochasticity is not sufficient
If stochasticity were itself the source of useful dynamics, it might seem that increasing it must improve the system.
DPSH assumes the opposite.
At too high a level of stochasticity:
structured causal influence
<<
random transitions.
The system then loses the capacity to maintain a metastable structure.
We may therefore assume three regimes:
Regime A – low stochasticity
σ ~ 0
Possible consequences:
rigid dynamics,
repeated trajectories,
dead states,
poor exploration.
Regime B – intermediate stochasticity
σ = σ*
Possible consequences:
exploration,
spontaneous transitions,
metastability,
sensitivity to weak evidence,
continued plasticity.
Regime C – high stochasticity
σ >> σ*
Possible consequences:
unstable representations,
excessive transitions,
loss of temporal structure,
poor prediction,
loss of percept persistence.
From this follows a testable prediction:
perceptual performance = G(σ)
need not be monotonic.
It may have a maximum at a non-zero value:
σ* > 0.
3.10 Stochasticity and symmetry breaking
Stochasticity acquires a further significance in a situation where several similarly stable states exist:
M_A ~ M_B.
An ambivalent sensory input, for example, may be compatible with both interpretations.
Without any asymmetry the system could theoretically remain in an unstable intermediate state.
A small fluctuation may, however, create:
activity_A = activity_B + ε.
If the recurrent dynamics amplifies this deviation:
ε
->
recurrent amplification
->
M_A,
a spontaneous breaking of symmetry occurs.
Stochasticity here does not determine the structure of the resulting percept.
It only initiates the selection among states whose structure already follows from the dynamics of the network and from sensory constraints.
This is an important distinction.
stochasticity != percept structure
but potentially:
stochasticity
->
selection among available perceptual structures.
3.11 Stochasticity and metastability
A stable attractor may lock the system in:
S -> M_A -> M_A -> M_A -> ...
Pure chaos, by contrast, does not provide sufficient persistence:
S1 -> S7 -> S3 -> S19 -> ...
DPSH seeks a regime between these extremes:
metastability.
In it the system maintains, for a certain time:
S(t) in M_A
but at the same time there is a non-zero probability:
P(M_A -> M_B) > 0.
Stochasticity may be one of the mechanisms that allow the current metastable state to be left.
The transition need not, however, be governed by stochasticity itself.
Its probability may depend on:
sensory evidence,
prediction error,
oscillatory phase,
recurrent state,
adaptation,
plasticity,
stochastic fluctuation.
In general:
P(M_A -> M_B) =
F(
evidence,
prediction,
phase,
history,
noise
).
3.12 Stochasticity and oscillation
Stochasticity by itself does not create temporal structure.
Local oscillations may, however, change the probability of a spontaneous spike over time.
For example:
P(spike_i, t) =
p0_i + A * f(φ(t)).
Spontaneous activity thereby ceases to be temporally homogeneous.
A spike is more probable in some phases than in others.
A combination arises:
stochasticity
+
temporal constraint.
This allows the network to preserve the exploratory character of stochasticity while at the same time creating structured temporal relations.
The working hypothesis of DPSH is therefore not:
noise -> percept.
It is:
stochastic exploration
+
oscillatory organization
+
recurrent selection
+
plasticity
+
sensory constraints
->
structured metastable dynamics.
3.13 Stochasticity and predictive processing
Predictive processing provides a mechanism by which spontaneous dynamics may be constrained by reality.
Let us assume that stochasticity creates a candidate state:
M_X.
This state generates a prediction:
P_X.
If the sensory evidence corresponds:
error(P_X, I) ~ 0,
the state may remain relatively stable.
If, however:
error(P_X, I) >> 0,
the prediction error may reduce its stability.
We thus obtain:
stochasticity
->
candidate states
->
prediction
->
comparison with environment
->
stabilization / destabilization.
The system may thus generate internal variability without being severed from external reality.
3.14 Perception as constrained stochasticity
From the preceding points a working interpretation follows:
Perceptual dynamics may arise as a constrained stochastic process in which local fluctuations enable exploration, while learned connectivity, recurrence, oscillations and sensory prediction constrain the probable trajectories of the system.
This gives rise to an important difference between:
random state
and
stochastic state.
A random state lacks a stable structure.
A stochastic dynamic state may be highly structured, even though its exact microscopic trajectory is not deterministic.
This leads to one of the important properties of DPSH:
The same percept need not correspond to the same configuration of individual spikes on every occurrence.
Two realizations:
E_A1
and
E_A2
may be microscopically different, but their global trajectory may belong to the same perceptual region:
E_A1 -> M_A
E_A2 -> M_A.
Perceptual identity would therefore exist at the macroscopic, not the microscopic level.
3.15 The first strong prediction
If spontaneous stochasticity has a functional role in the formation of internal states, there must be an experimentally observable difference between:
stochastic network
and
otherwise equivalent deterministic network.
It is not sufficient, however, to compare only the resulting accuracy.
It is necessary to track:
number of metastable states,
state lifetime,
transition probability,
state-space coverage,
trajectory diversity,
robustness,
generalization,
recovery after perturbation,
spontaneous/evoked state similarity.
DPSH predicts that suitable non-zero stochasticity will increase at least some of these properties without a loss of perceptual separability.
3.16 Experiment S1 – sweep of spontaneous activity
The first basic experiment in Cognia will vary the baseline stochasticity:
σ = {
0,
0.001,
0.005,
0.01,
0.05,
0.1,
...
}.
The exact values will depend on the implemented neuronal model.
For each value we will measure:
firing rate,
number of active neurons,
state-space coverage,
number of metastable clusters,
cluster lifetime,
transition entropy,
response to sensory perturbation,
perceptual task performance.
The hypothesis assumes the existence of a non-zero region in which the
dynamic structure of the system attains better properties than at
σ = 0.
3.17 Experiment S2 – spontaneous activity after experience
We first expose the network to a structured environment:
A, B, C, D ...
and allow plasticity.
We subsequently remove the external input:
I_external = 0.
We compare the spontaneous dynamics:
before learning
and
after learning.
If learning has changed the internal dynamic space, we expect:
spontaneous_before
!=
spontaneous_after.
The stronger prediction is that after learning, spontaneous activity will more frequently visit regions of the state space similar to the states evoked by the learned sensory structures.
One can measure, for example:
D(spontaneous_state, evoked_state).
This experiment directly tests the idea that experience changes the geometry of the internal dynamics.
3.18 Experiment S3 – switching off spontaneous firing
After the network has been trained we create two identical copies:
Network A:
baseline firing ON
Network B:
baseline firing OFF.
The external input will subsequently be removed for a certain time.
We then restore an incomplete or ambivalent input.
We will test whether the previous internal state influences the subsequent interpretation and whether its preservation depends on spontaneous activity.
If:
spontaneous firing OFF
has no measurable influence on:
state persistence,
later interpretation,
state-space structure,
learning,
then the hypothesis about its fundamental functional role will be weakened.
3.19 Experiment S4 – separating noise from firing rate
One of the greatest experimental risks is the confusion of:
stochasticity
with mere:
increased activity.
It is therefore necessary to create a control in which two networks have a similar:
mean firing rate,
spike count,
energy/activity level,
but different temporal stochasticity.
For example:
Network A:
stochastic spike timing
Network B:
matched firing rate,
deterministic or replayed timing.
If their macroscopic dynamics differ significantly, it will be possible to argue that the relevant quantity is not merely the amount of activity, but its stochastic temporal organization.
3.20 Experiment S5 – frozen noise
Another important control is the so-called frozen stochastic sequence.
We first generate a particular sequence of random events:
R = {r1, r2, ..., rn}.
We then use it identically across repeated runs.
We compare:
fresh stochasticity
against:
identical replayed stochasticity.
Both variants may have:
same spike probability,
same distribution,
same expected firing rate,
but only the first generates new microscopic trajectories.
In this way it is possible to distinguish whether what matters for the system is merely the presence of a noise-like signal, or genuine ongoing exploration of new trajectories.
3.21 Falsification criteria
The strong version of the hypothesis about spontaneous stochastic activity will be weakened if experiments show that:
σ = 0creates equally rich or richer metastable dynamics,- increasing stochasticity only degrades the representation,
- spontaneous firing after the removal of input in no way contributes to the persistence of the internal state,
- learned experience does not change the structure of spontaneous activity,
- stochastic timing has no effect other than a corresponding increase of the firing rate,
- spontaneous activity does not contribute to plasticity or to future processing,
- all the assumed effects can be explained by a simpler deterministic mechanism.
In such a case stochasticity must be removed from the central mechanism of DPSH or reclassified as a secondary biological property.
3.22 Chapter research hypothesis
We formulate the partial hypothesis H2:
H2 – Functional Stochasticity Hypothesis
Non-zero spontaneous stochastic activity in a globally clockless recurrent network enables exploration of the internal state space and, in interaction with recurrence, plasticity and temporal organization, contributes to the formation, maintenance of and transitions between metastable population states. There exists a region of stochasticity in which these dynamic properties are more pronounced than in the corresponding deterministic system.
The hypothesis does not say that stochasticity creates the percept by itself.
It claims only that it may be one of the causally significant conditions of the dynamics out of which the percept arises.
Its validity must be assessed independently of the other parts of DPSH.
4. Endogenous oscillators as a temporal structure
4.1 Time inside the network
The Dynamic Perceptual State Hypothesis distinguishes between two fundamentally different ways of working with time.
The first possibility is time imposed on the system from outside:
global clock
->
update all units
->
next global state.
The second possibility is a temporal structure arising inside the network itself:
local dynamics
->
oscillatory activity
->
phase-dependent modulation
->
temporally structured interaction.
DPSH proceeds from the second principle.
A network need not have a global mechanism determining when a computation is to take place. It may nevertheless contain a very rich temporal organization.
Oscillations on this account are part of the state of the system.
They do not determine the moments of global update, but they change the conditions under which individual local events acquire significance.
4.2 A global clock is not a local oscillator
It is necessary to separate consistently:
global update clock
from:
endogenous neural oscillator.
A global clock says:
"all units now perform the next step."
An endogenous oscillator says only:
"a certain time-varying signal is currently present in this part of the system."
A neuron may respond to this signal, but it need not wait for its next period in order to process another event.
Therefore:
oscillator != scheduler.
The literature review showed that this difference also has technical precedents: there exist event-driven spiking systems without a global clock that nevertheless use local oscillators.
4.3 Oscillation as part of the neuronal state
For a local oscillator Ok we may introduce a state:
Ok(t) = {
phase,
frequency,
amplitude
}.
A simple periodic model may take the form:
Ok(t) = Ak * sin(ωk*t + φk).
This equation is not, however, the substance of the hypothesis.
What matters is that Ok(t) enters into the dynamics of the neurons.
For example:
P(spike_i, t) =
F(
si(t),
Ii(t),
Ri(t),
Ok(t),
ξi(t)
).
The phase may therefore change the probability that a neuron emits a spike at a given moment.
The same neuron with the same input therefore need not respond in the same way in different phases:
response(input, phase_A)
!=
response(input, phase_B).
4.4 Phase as a dynamic context
In the classical rate-based view, the activity of a neuron may be described primarily by the quantity:
firing_rate.
In DPSH, however, the significance of a neuron may be extended by its relation to the local phase:
neural_event =
{
source,
time,
phase_context
}.
Two spikes with the same source and the same approximate firing rate may have a different functional significance if they occurred in different phases.
The literature review showed an experimental precedent for this idea: phase-of-firing coding may carry additional information beyond the number of spikes itself.
Time thereby does not become merely a physical coordinate.
It becomes a potential part of the neuronal representation.
4.5 Oscillatory modulation of probability
For a stochastic neuron one may introduce an instantaneous spiking intensity:
λ_i(t).
This may be influenced, for example, as:
λ_i(t) =
g(
baseline_i,
sensory_i(t),
recurrent_i(t),
phase_i(t),
stochastic_state_i(t)
).
An oscillator therefore need not generate a spike directly.
It may only periodically change:
excitability,
threshold,
synaptic gain,
spike probability,
plasticity sensitivity.
This is important.
Temporal structure may arise as a modulation of the probability of events, not as their deterministic scheduling.
4.6 Multiple local oscillators
A network need not contain a single dominant rhythm.
Let us consider a set:
O = {O1, O2, ..., Om}.
Each may have its own:
frequency,
phase,
amplitude,
spatial influence,
coupling.
A neuron Ni may be influenced by a subset:
O_i = {O2, O5, O8}.
Then:
P(spike_i,t) =
F(
...,
φ2(t),
φ5(t),
φ8(t)
).
A combinatorially very rich temporal structure thereby arises.
The same neuron may find itself in a different dynamic context depending on what the relative phases of several local rhythms currently are.
4.7 Relative phase
For two oscillations let us define:
Δφ_ij = φ_i - φ_j.
DPSH assumes that some functional properties of the network may depend rather on:
Δφ
than on the absolute phase of an individual oscillator.
For example:
communication_efficiency =
F(Δφ).
If two populations are in a suitable phase relation:
Δφ ~ Δφ_optimal,
the efficiency of transmission may be high.
At a different relative phase:
Δφ ~ Δφ_nonoptimal
transmission may be weakened.
This gives rise to the possibility of time-varying functional connectivity without a change of the anatomical synapses.
4.8 Dynamic routing
A fixed network may have the topology:
A -> B
A -> C
A -> D.
This need not mean, however, that A communicates with all targets
equally effectively at every moment.
If:
B is receptive at phase φ1
C is receptive at phase φ2
D is receptive at phase φ3,
then the same spike from A may have a different effect:
A -> B strong
A -> C weak
A -> D none.
At another moment the distribution may be the opposite.
The literature review showed that the phase of a local oscillation may indeed change the efficacy of an incoming spike volley, and that interareal synchronization may be related to selective effective connectivity.
DPSH therefore works with the hypothesis:
Relative phase may function as a dynamic routing mechanism.
Such routing is not governed by a central scheduler.
It arises from the instantaneous dynamic state of the populations.
4.9 Oscillation and stochasticity
Spontaneous stochasticity generates variability:
possible spike
->
possible trajectory.
Oscillations may organize this variability temporally.
Instead of:
random event probability = constant
it may hold that:
random event probability = phase-dependent.
For example:
P(spike_i,t)
=
p0_i + A_i * f(φ(t)).
An interesting combination thereby arises:
stochasticity
->
exploration
while:
oscillatory phase
->
temporal constraint.
DPSH therefore does not regard stochasticity and oscillations as opposites.
On the contrary, they may form a single mechanism:
variability
+
temporal organization
->
structured stochastic dynamics.
4.10 An oscillation need not have its own specialized cell
The original architectural intuition may tempt one towards the model:
oscillator cell
->
neuron population.
This is a legitimate implementation possibility.
The literature review, however, showed an important precedent: population oscillations may emerge from recurrent stochastic spiking units without the individual neurons themselves being intrinsic oscillators.
DPSH therefore distinguishes two architectures.
Explicit oscillator
oscillator unit
->
local population.
Emergent oscillation
recurrent population
->
population rhythm
->
modulation of population.
The hypothesis does not commit itself to the claim that conscious dynamics needs a special type of oscillator cell.
More substantial may be the existence of:
endogenous local oscillatory dynamics.
4.11 Oscillation as an emergent macrostate
This leads to a stronger possibility.
Oscillations need not be merely an input into the perceptual dynamics.
They may at the same time be its result.
That is:
recurrent connectivity
+
stochastic spiking
->
oscillatory population state.
And this state subsequently influences, in turn:
spike timing,
communication,
plasticity.
A closed loop arises:
local connectivity
|
v
population rhythm
|
v
spike timing
|
v
plasticity
|
v
modified connectivity
|
+--------------+
|
v
population rhythm.
This cycle constitutes a candidate mechanism of self-organization.
4.12 Oscillation and synaptic delays
In a globally clockless network, delays are of fundamental significance.
For a synapse:
i -> j
we define:
d_ij.
A spike emitted at time:
t_i
arrives at:
t_i + d_ij.
If the target neuron is modulated by an oscillation, the effect of the spike also depends on:
φ_j(t_i + d_ij).
The effect of the synapse is therefore not merely a function of:
weight_ij.
It is a function of:
effect_ij =
F(
weight_ij,
delay_ij,
arrival_phase,
local_state_j
).
Two synapses with the same weight may have a very different functional effect if they have different delays.
4.13 The temporal topology of the network
From the preceding point it follows that the network does not have only a spatial topology:
who is connected to whom.
It also has a temporal topology:
when can influence whom.
This topology is given by the combination of:
synaptic delays,
oscillatory phases,
refractory periods,
adaptation,
stochastic spike timing.
DPSH therefore regards the temporal structure of the network as equally important as the connectivity itself.
4.14 Oscillation and STDP
If plasticity depends on the relative timing of pre- and postsynaptic spikes:
Δt = t_post - t_pre,
then oscillations may indirectly change learning by structuring the timing of spikes.
The mechanism:
oscillatory phase
->
spike probability
->
spike timing
->
STDP
->
synaptic structure.
The literature review showed that the combination of oscillatory input, spike timing, delays and STDP may indeed select connectivity and create distributed attractor structures.
This means that temporal organization need not merely modulate an already learned network.
It may actively participate in how the network learns.
4.15 The closed phase-plasticity loop
DPSH therefore assumes the possibility of the following loop:
phase relations
->
spike timing
->
STDP
->
synaptic weights and effective delays
->
population dynamics
->
new phase relations.
Formally:
Φ(t)
->
E(t)
->
W(t + dt)
->
S(t + dt)
->
Φ(t + dt).
The network may thereby gradually create temporally compatible dynamic structures.
This is important for the formation of metastable states.
4.16 Resonance
If some populations respond preferentially to a certain temporal structure, resonance may appear.
For example:
input frequency ≈ local preferred frequency
may lead to:
stronger propagation,
increased synchronization,
higher spike probability,
stronger plasticity.
A different input:
input frequency far from preferred frequency
may have a smaller effect.
This allows selective processing without the need for an explicit logical gate.
4.17 Interference
Several oscillations may create a combined temporal structure.
For two oscillations:
O1(t)
O2(t)
their combination may create moments of:
constructive alignment
and:
destructive alignment.
A neuron may respond approximately to:
O_total(t) = O1(t) + O2(t).
DPSH does not say that neuronal interference is identical with quantum interference.
It uses only the general dynamic principle:
Several periodic or quasiperiodic influences may jointly create time-varying regions of increased and decreased excitability.
Such interferences may significantly extend the number of dynamic configurations available to the system.
4.18 Cross-frequency coupling
Different frequencies need not operate independently.
For example, a slow rhythm may modulate the amplitude or the efficacy of faster activity:
slow phase
->
fast oscillation amplitude.
In general:
A_fast(t) =
F(φ_slow(t)).
A hierarchical temporal structure thereby arises.
Slow rhythms may define broader dynamic windows, while faster rhythms may organize the finer timing of events.
DPSH regards this property as potentially relevant for the hierarchical arrangement of the perceptual state.
4.19 Oscillation and non-commutativity
If the effect of an event depends on the phase, then the order of events is naturally non-commutative.
For example:
spike A at phase φ1
spike B at phase φ2
may create the state:
S_AB.
The opposite order:
spike B at phase φ1
spike A at phase φ2
may create:
S_BA.
In general:
S_AB != S_BA.
Oscillations therefore provide one of the mechanisms that convert temporal order into different state-space trajectories.
4.20 Oscillation and symmetry breaking
Let us assume two competing populations:
A
B.
Both have similar support:
support(A) ≈ support(B).
If, however, at a given moment:
phase_A favorable
and:
phase_B unfavorable,
the same input may cause:
response_A > response_B.
A small temporal asymmetry may be amplified by recurrence:
phase difference
->
small activity difference
->
recurrent amplification
->
symmetry breaking
->
selected metastable state.
The relative phase may thereby influence which of several possible perceptual states will be realized.
4.21 Oscillation and metastable states
DPSH does not assume that oscillations are to create one permanently synchronized state.
On the contrary.
An oscillatory structure may enable:
temporary coherence
->
state formation
->
phase drift
->
reduced coherence
->
transition.
That is:
M_A
->
phase reorganization
->
transition
->
M_B.
Oscillations may thus contribute simultaneously to:
stabilization
as well as:
destabilization.
This is naturally compatible with metastability.
4.22 The percept as a phase-organized macrostate
The stronger working hypothesis of this chapter is:
A perceptually relevant metastable state need not be defined solely by the set of active neurons or by their firing rates. Part of its identity may be contained in the relative temporal and phase relations between neuronal populations.
Two realizations may then have:
similar firing rates
but:
different phase geometry.
And therefore correspond to different dynamic states:
M_A != M_B.
This is one of the most important testable parts of DPSH.
4.23 Phase geometry
For a set of local oscillations the state can be described in simplified terms by a vector:
Φ(t) =
(
φ1(t),
φ2(t),
...,
φm(t)
).
The relative phase relations then determine a point in the so-called phase space.
DPSH examines the possibility that some perceptual states correspond not only to regions of the neuronal state space:
S(t) in M_A,
but at the same time to regions of the phase geometry:
Φ(t) in P_A.
A percept may therefore be defined by the combination of:
neuronal state
+
temporal organization.
Schematically:
Percept_A =
M_A × P_A.
This is not a final mathematical definition.
It is a working model for experimental testing.
4.24 Oscillation and the continuity of the percept
If a percept is composed of many local dynamic processes, their temporal coordination may enable temporary coherence without a central controller.
Individual neurons may enter and leave activity.
Nevertheless a relational structure may persist:
neuron set changes
but:
phase organization persists.
This offers a candidate mechanism for how a global percept may remain relatively stable even though its microscopic neuronal substrate is constantly changing configuration.
4.25 Explicit versus emergent oscillator in Cognia
Cognia should experimentally support at least two variants.
Variant A – explicit oscillator
A separate unit:
oscillator O {
frequency
phase
amplitude
}
which generates a local modulatory signal.
Variant B – emergent oscillation
A recurrent microcircuit:
excitatory population
+
inhibitory population
+
delays
->
emergent rhythm.
These two variants must be separated experimentally.
If both produce the same relevant effect, then:
oscillator cell
is not a necessary component of the hypothesis.
What may be necessary is only:
local oscillatory dynamics.
This also corresponds to the conclusion of the literature review that explicit oscillator cells are not a necessary precondition of a population oscillation.
4.26 Experiment O1 – phase scrambling
This is the main experiment of the chapter.
We first create a network that exhibits stable or metastable perceptual representations.
We then compare:
condition A:
phase relations intact
and:
condition B:
phase relations scrambled.
We must preserve as far as possible:
mean firing rate,
spike count,
sensory input,
topology,
synaptic weights,
network size.
We manipulate primarily:
relative timing structure.
We measure:
state separability,
metastable lifetime,
decoding accuracy,
transition entropy,
percept persistence,
behavioral performance.
The strong prediction:
Q_intact > Q_scrambled
even at:
firing_rate_intact ≈ firing_rate_scrambled.
This experiment was also identified in the literature review as the main causal test of the hypothesis.
4.27 Experiment O2 – phase jitter
Phase scrambling is a coarse manipulation.
We therefore introduce a gradual jitter:
jitter = {
0 ms,
1 ms,
2 ms,
5 ms,
10 ms,
20 ms,
...
}.
We track whether the quality of the dynamic state declines:
Q(jitter).
If there is a particular temporal scale at which the representation begins to collapse, we obtain an estimate of the temporal precision relevant for the given network.
4.28 Experiment O3 – frequency shift
While preserving the approximate amplitude of the oscillation, we will vary:
frequency.
For example:
f1,
f2,
f3,
...
We track:
state formation,
learning speed,
state stability,
transition probability.
If preferred dynamic frequencies exist, they need not be arbitrary.
They may be the result of the interaction of:
synaptic delays,
refractory periods,
STDP windows,
recurrent topology.
4.29 Experiment O4 – explicit versus emergent oscillation
We compare:
A:
explicit oscillator units
B:
emergent oscillatory microcircuits
C:
matched nonoscillatory network.
We control approximately:
firing rate,
network size,
input,
capacity.
The main question:
Is a particular type of oscillator important for the formation of metastable perceptual states, or only the existence of a functional temporal organization?
If:
A ≈ B > C,
then the hypothesis will support the more general principle:
oscillatory dynamics
instead of:
oscillator cells.
4.30 Experiment O5 – relative phase between populations
We create two functionally connected populations:
A
B.
We will systematically vary:
Δφ_AB.
For example:
0°
45°
90°
135°
180°.
With the same anatomical connection we measure:
effective transmission,
spike propagation,
downstream state changes.
In this way it is possible to test directly:
effective_connectivity =
F(relative_phase).
4.31 Experiment O6 – phase and the ambivalent percept
The network receives an ambivalent input supporting two states:
M_A
M_B.
Before the presentation of the input we set different relative phases of the local oscillations.
If:
identical sensory input
+
different initial phase configuration
leads systematically to:
different perceptual state selection,
this will be evidence that the internal temporal state of the network influences the interpretation of the input.
This experiment directly connects oscillations with hysteresis and symmetry breaking.
4.32 Experiment O7 – phase and plasticity
We compare:
phase structured + STDP
phase scrambled + STDP
phase structured + plasticity OFF
phase structured + rate-based plasticity.
We measure:
learned state geometry,
state separability,
attractor/metastable structure,
generalization.
In this way it is possible to determine whether phase organization merely modulates activity in the short term, or genuinely shapes the long-term structure of the network.
4.33 Falsification criteria
The strong oscillatory hypothesis will be weakened if:
- phase scrambling with a preserved firing rate does not influence the relevant state-space dynamics,
- phase jitter has no systematic effect,
- relative phase does not change effective connectivity,
- oscillations neither increase nor change the formation of metastable states,
- a nonoscillatory control network creates the same representations with the same robustness,
- phase-dependent plasticity provides no advantage over simpler learning rules,
- all the effects of oscillations can be explained solely by a change of firing rate,
- the initial phase state has no causal influence on the interpretation of an ambivalent input.
In such a case, oscillations must be removed from the central part of DPSH or regarded merely as one possible implementation of a more general temporal dynamics.
4.34 Chapter research hypothesis
We formulate the partial hypothesis H3:
H3 – Endogenous Temporal Organization Hypothesis
In a globally clockless recurrent spiking network, local endogenous oscillations may organize otherwise stochastic spike events through phase-dependent excitability, communication and plasticity. Relative phase relations may thereby constitute a functionally significant state variable that contributes to the formation, stabilization of and transitions between metastable population states.
The stronger falsifiable prediction is:
If relative phase genuinely carries causally relevant information, then its specific disruption while preserving the mean firing rate, spike count, connectivity and sensory input must selectively impair at least some properties of the metastable perceptual state.
This hypothesis does not claim:
oscillation = percept
nor:
oscillation = consciousness.
It claims only:
oscillatory temporal organization
->
functionally relevant structure
->
contribution to perceptual dynamics.
Only the following chapters must show whether this organization genuinely leads to the formation of an integrated metastable perceptual state.
5. Non-commutative neuronal dynamics
5.1 Why order matters
One of the main consequences of globally clockless, stateful and temporally structured neuronal dynamics is the fact that the order of events may change the resulting state of the system.
If we have two events or two transformations:
A
B
then in general it need not hold that:
A(B(S)) = B(A(S)).
On the contrary, we expect:
A(B(S)) != B(A(S)).
Such dynamics is non-commutative.
This means that a neuronal system cannot be fully described merely by the set of events that occurred.
It is also necessary to know:
their order,
their relative timing,
the state of the system at the moment of their arrival,
the phase of the local oscillations,
the synaptic history,
the current plasticity.
History is therefore not merely a record of the past.
It is an active part of the present state of the system.
5.2 Commutative and non-commutative transformations
Let us consider a simple state:
S0.
Two events act upon it:
A
B.
In a commutative system:
S_AB = B(A(S0))
and:
S_BA = A(B(S0))
where:
S_AB = S_BA.
The order of the events is of no significance.
In a non-commutative system:
S_AB != S_BA.
The same pair of events therefore creates two different states depending on which of them arrived first.
This is important for systems that are to represent a world evolving in time.
The sequence:
door opens
person enters
is not equivalent to the sequence:
person enters
door opens.
It contains a different causal structure.
5.3 Non-commutativity as a consequence of the internal state
If the response of a neuron is a function of its current state:
response = F(input, state),
then the first event changes the state upon which the second acts.
That is:
S1 = F_A(S0)
and then:
S2 = F_B(S1).
In the opposite order:
S1' = F_B(S0)
and:
S2' = F_A(S1').
Therefore in general:
S2 != S2'.
Non-commutativity therefore need not be a specially implemented property.
It arises naturally in every system in which:
the current state depends on the past
and
events change this state.
5.4 The refractory period as a simple example
Let us imagine a neuron with a refractory period.
Spike A arrives at time:
t0
and activates the neuron.
Shortly afterwards spike B arrives:
t0 + dt.
If the neuron is still refractory:
B -> weak/no response.
In the opposite order:
B first
A second
A may be suppressed.
So:
response(A -> B)
!=
response(B -> A).
The order of the inputs changes their functional effect.
5.5 Synaptic integration
A further source of non-commutativity is temporal integration.
A neuron may integrate inputs within a certain window:
τ.
Two spikes:
A at t1
B at t2
may jointly cross the threshold.
If, however, one of them is inhibitory:
A = excitation
B = inhibition,
then:
excitation -> inhibition
need not have the same effect as:
inhibition -> excitation.
The first sequence may evoke a spike before the inhibition arrives.
The second may prevent its occurrence.
The resulting network trajectory may then diverge fundamentally.
5.6 Oscillatory phase and order
In the previous chapter we showed that the effect of a spike may depend on the local phase.
Then two identical events:
A
B
may act in different phases:
A at φ1
B at φ2.
In the opposite order:
B at φ1
A at φ2.
This need not be equivalent.
In general:
effect(A, φ1) + effect(B, φ2)
!=
effect(B, φ1) + effect(A, φ2).
An oscillatory structure may therefore markedly strengthen the non-commutative character of the network.
5.7 STDP as an explicit non-commutative mechanism
Spike-timing-dependent plasticity provides a very direct example.
If the presynaptic neuron spikes before the postsynaptic one:
pre -> post
a different synaptic change may occur than in the case of:
post -> pre.
Formally:
Δw(pre, post)
!=
Δw(post, pre).
The order of events thereby does not change only the immediate activity.
It also changes the future structure of the network.
We obtain two degrees of non-commutativity:
order
->
different current state
and at the same time:
order
->
different future dynamics.
5.8 Non-commutativity and learning
If a sequence of events changes the weights:
A -> B
->
W_AB
while:
B -> A
->
W_BA,
and:
W_AB != W_BA,
then experience changes the geometry of the future state space according to the causal history.
The network therefore does not learn only:
what occurred,
but also:
in what order it occurred.
This is fundamental for:
sequences,
causality,
prediction,
motor programs,
language,
spatiotemporal relations.
5.9 Non-commutativity and predictive processing
Predictive processing naturally works with temporal structure.
If the system expects:
A -> B,
then the sequence:
A -> B
may be readily predictable.
The sequence:
B -> A
may create a prediction error.
This means that the internal model need not represent only the probability of individual events:
P(A),
P(B),
but also the conditional relations:
P(B | A)
and:
P(A | B).
In general:
P(B | A) != P(A | B).
Temporal asymmetry thus becomes part of the internal model of the world.
5.10 Causality
Non-commutativity is closely connected with causality.
If:
A causes B,
then the sequence:
A -> B
has a different meaning than:
B -> A.
DPSH assumes that the internal perceptual state must be sensitive not only to present correlations, but also to the directionality of interactions.
The internal representation of the world therefore need not be merely spatial.
It must also be causal-temporal.
5.11 The state-space trajectory
In a commutative system the resulting state may be predominantly a function of the set of inputs:
S_final = F({A, B, C}).
In a non-commutative system it is more appropriate to write:
S_final = F(A -> B -> C).
The sequence:
A -> B -> C
leads along the trajectory:
S0 -> S1 -> S2 -> S3.
The sequence:
C -> B -> A
may lead:
S0 -> S1' -> S2' -> S3'.
DPSH therefore regards the trajectory through the state space as more important than the final point itself.
5.12 Path dependence
Non-commutativity leads to a more general property:
path dependence.
The same final sensory input may be reached by different paths:
path A
->
input X
and:
path B
->
input X.
If the system maintains a history:
S_A != S_B,
then:
F(S_A, X)
!=
F(S_B, X).
A mechanism thereby arises by which previous experience influences present perception.
5.13 Non-commutativity and hysteresis
Hysteresis can be understood as a macroscopic consequence of path dependence.
Upon the change:
A -> B
the system may remain in the state M_A up to a certain threshold.
In the opposite direction:
B -> A
the threshold may be different.
That is:
threshold(A -> B)
!=
threshold(B -> A).
This shows that the present state cannot be determined solely from the current value of the input.
It also depends on the path the system has traversed.
5.14 Non-commutativity and the percept
If a percept corresponds to a metastable dynamic state:
M,
then the path by which the network entered M may influence its fine
internal structure.
We may therefore have:
M_A^path1
and:
M_A^path2
which correspond to a similar macroscopic percept, but differ in:
phase relations,
synaptic state,
local activation,
transition probabilities.
This offers an important possibility:
Two subjectively similar percepts need not be microscopically identical.
Their identity may exist at a higher level of dynamic organization.
5.15 Order as information
In such a system the order itself becomes a carrier of information.
For example:
A -> B -> C
may represent a different meaning than:
A -> C -> B.
Even though the set of events is the same:
{A, B, C}.
This is important for language.
The sequence of words:
the dog bites the man
is not equivalent to:
the man bites the dog
even in the case where very similar conceptual representations were activated.
The same principle holds for:
motor control,
music,
spatial events,
social interactions,
causal inference.
5.16 Non-commutativity and temporal coding
If information depends on order, it cannot be fully captured by the average firing rate alone.
Two sequences may have:
same neurons,
same spike count,
same average firing rate,
but a different order:
A -> B -> C
versus:
C -> B -> A.
DPSH assumes that such sequences may create different internal states.
This provides a very strong experimental design, because the statistical activity can be preserved and only the timing manipulated.
5.17 Non-commutativity and symmetry breaking
Let us imagine an ambivalent state:
M_A ~ M_B.
A small event X may shift the system towards M_A.
A subsequent event Y then acts upon an already changed state.
The sequence:
X -> Y
may therefore stabilize M_A.
Conversely:
Y -> X
may stabilize M_B.
The order of local fluctuations may thus decide which global interpretation will ultimately be realized.
Non-commutativity may therefore function as one of the microscopic mechanisms of spontaneous symmetry breaking.
5.18 Non-commutativity and metastability
A metastable state has a finite lifetime.
The probability of leaving it may depend on the sequence of incoming events:
P(M_A -> M_B | X -> Y)
!=
P(M_A -> M_B | Y -> X).
This means that the transition graph of the system is not merely a function of the set of stimulation events.
It is a function of temporally ordered sequences.
5.19 Composition of transformations
For a formal description, transformations may be assigned to individual events or modules:
T_A
T_B
T_C.
The evolution of the system:
S' = T_C T_B T_A S.
If the transformations do not commute:
[T_A, T_B] != 0,
where we define the commutator as:
[T_A, T_B] =
T_A T_B - T_B T_A,
then the order of their application changes the state.
This notation is useful as a mathematical inspiration.
DPSH does not thereby claim that the neuronal system is a quantum system.
Non-commutativity here arises from classical nonlinear, stateful and history-dependent dynamics.
5.20 Why a quantum hypothesis is not needed
The similarity with non-commutative operators in quantum mechanics may be intuitively interesting.
It is not, however, necessary to assume:
quantum brain
nor:
quantum computation.
A classical dynamic system with:
nonlinearity,
memory,
delays,
adaptation,
recurrence
may be strongly non-commutative.
DPSH uses non-commutativity as a general mathematical principle:
The result of a sequence of transformations depends on their order.
It thereby avoids the unsupported transfer of quantum mechanisms into neuronal dynamics.
5.21 Non-commutativity and internal experience
If the present state depends on the entire trajectory:
S(t) = F(history),
then the experience of the system is not merely an archive of past data.
The past is physically encoded in the present:
synaptic state,
membrane state,
phase state,
adaptation,
network trajectory.
This leads to an important principle:
history
->
current state
->
interpretation of future input.
Internal experience may therefore be understood as a state trace of the system's previous interaction with the world.
5.22 Non-commutativity and intuition
This principle may later also be related to intuitive decision-making.
Over a long experience a network may traverse a great number of trajectories:
experience
->
plasticity
->
learned state-space geometry.
A new input may then very rapidly move the system into the region:
M_A
without the need to reconstruct explicitly all the previous causal steps.
The decision:
action_A
may thus be the result of the entire learned dynamics, even though the system has no globally available explicit representation of:
"why I chose A".
In this sense, intuitive decision-making may be a macroscopic consequence of historically shaped non-commutative dynamics.
5.23 Non-commutativity and the Perceptual Manifold
If transitions are non-commutative, the Perceptual Manifold cannot be understood merely as a set of points.
It is necessary to include:
states
+
directed transitions
+
transition histories.
Formally it may be more appropriate to write:
M = (S, E)
where:
S = set of perceptual states
and:
E = directed transitions.
The transition:
M_A -> M_B
need not be equivalent to:
M_B -> M_A.
The Perceptual Manifold thereby acquires a directional structure.
5.24 Temporal geometry
If different paths between states have different consequences, then the distance between two states need not be symmetric in a purely functional sense.
For example:
cost(M_A -> M_B)
!=
cost(M_B -> M_A).
Likewise:
transition_probability(M_A -> M_B)
!=
transition_probability(M_B -> M_A).
The Perceptual Manifold may therefore have not only a geometry of states, but also a dynamic orientation.
5.25 Experiment N1 – order reversal
The basic experiment will use two events:
A
B.
We compare:
A -> B
and:
B -> A.
We control:
same inputs,
same duration,
same number of events,
same approximate firing rate,
same initial state distribution.
We measure:
D(S_AB, S_BA),
trajectory divergence,
state separability,
later behavioral effect.
If:
D(S_AB, S_BA) ~ 0
for all relevant conditions, the strong version of the order-dependence hypothesis will be weakened.
5.26 Experiment N2 – timing continuum
The order will remain:
A -> B,
but we will vary:
Δt = t_B - t_A.
For example:
-50 ms
-20 ms
-10 ms
-5 ms
0 ms
5 ms
10 ms
20 ms
50 ms.
We thereby obtain a function:
Q(Δt).
If timing genuinely influences the dynamic state, there should be a
structured dependence on Δt.
5.27 Experiment N3 – matched-rate sequence test
We create two sequences:
sequence 1:
A -> B -> C
sequence 2:
C -> B -> A.
We ensure that the following are as similar as possible:
neuron participation,
spike count,
mean firing rates,
stimulus energy.
We manipulate only:
temporal order.
If a decoder can reliably distinguish the resulting internal states even after the sequence has ended, we have evidence for an order-dependent representation.
5.28 Experiment N4 – order-dependent perception
We use:
A -> B -> ambiguous X
and:
B -> A -> ambiguous X.
X itself is identical.
If:
P(Y_A | A -> B -> X)
!=
P(Y_A | B -> A -> X),
then the temporal history changes the subsequent interpretation of the same stimulus.
This connects non-commutativity directly with perceptual function.
5.29 Experiment N5 – order-dependent learning
We will train the network in two regimes:
training 1:
A -> B
training 2:
B -> A.
After training we compare:
W_AB
and:
W_BA,
but also:
spontaneous dynamics,
metastable state geometry,
response to incomplete inputs.
In this way we find out whether the order of experiences changes not only the local synapses, but the global state space.
5.30 Experiment N6 – commutative control
It is important to create a control model that will be deliberately more commutative.
For example:
aggregate all spikes in window
->
calculate rate
->
update state.
Such a model may preserve:
spike count,
average activity,
input identity,
but remove part of the timing information.
A comparison with the event-driven DPSH network will show whether non-commutative temporal structure provides a functional advantage.
5.31 Measuring the degree of non-commutativity
For two transformations we may define a simple empirical measure:
C(A,B,S) =
D(
T_B(T_A(S)),
T_A(T_B(S))
).
If:
C ~ 0,
the transformations are approximately commutative in the given state.
If:
C >> 0,
the order has a marked effect.
What is important is that:
C
may depend on the state itself:
C = C(A,B,S).
The network therefore need not be globally non-commutative to the same degree.
Non-commutativity may be a local property of certain regions of the state space.
5.32 Non-commutativity as an experimental variable
This makes it possible to create a map:
state-space region
->
degree of noncommutativity.
It may turn out, for example, that:
stable trivial states
->
low C
while:
metastable perceptual regions
->
high C.
If such a relation existed, it would be very interesting.
It would mean that rich perceptual dynamics is associated with a higher sensitivity to the order of events.
5.33 Falsification criteria
The strong hypothesis of non-commutative perceptual dynamics will be weakened if:
- reversing the order of events while preserving the other statistics does not change the internal state,
- a change of
Δthas no systematic effect, - history-dependent differences can be fully explained by a simple explicit memory variable alone,
- timing-sensitive plasticity provides no result different from rate-based learning,
- an ambivalent percept is not influenced by the previous sequence,
- the Perceptual Manifold is equally well described by an unoriented, history-independent representation.
In such a case non-commutativity would not be a central mechanism of DPSH, but merely a local property of certain neuronal processes.
5.34 Chapter research hypothesis
We formulate the partial hypothesis H4:
H4 – Non-Commutative Neural Dynamics Hypothesis
In a globally clockless stateful neuronal network, the relative order and timing of events is a functionally significant part of the computation. Therefore two sequences containing the same local events may lead to different neuronal trajectories, synaptic changes and subsequent perceptual states if their causal and temporal order differs.
The stronger prediction reads:
If non-commutativity is an important property of the formation of a perceptual state, then order reversal and timing perturbation must change the subsequent metastable representation even while preserving the identity of the inputs, their number and the approximate population activity.
The hypothesis therefore does not assume:
noncommutativity = consciousness.
It claims:
temporal order
->
different state trajectory
->
different internal representation.
The history of the system thereby becomes directly part of its present perceptual dynamics.
6. Self-organization and spontaneous symmetry breaking
6.1 From local activity to global state
The Dynamic Perceptual State Hypothesis assumes that a coherent perceptual state need not be created by a central mechanism that explicitly assembles the individual parts of the representation into a single whole.
It may arise as a macroscopic consequence of local interactions.
The basic scheme is:
local activity
+
recurrence
+
stochastic fluctuations
+
temporal organization
+
inhibition / competition
->
self-organized population state.
The network therefore need not contain a unit of the type:
select_percept(A).
The selection may be the result of the dynamics of the whole population.
6.2 What self-organization means
By self-organization we understand the emergence of a structure that is not explicitly written into a single central controlling component.
Every local unit responds only to a limited amount of information:
local state,
incoming spikes,
modulatory signals,
oscillatory phase,
synaptic history.
Nevertheless their joint interaction may create a global property:
Ψ(S).
This global property need not be available to any individual neuron.
An example may be:
population coherence,
phase organization,
attractor occupancy,
metastable state identity,
perceptual interpretation.
DPSH assumes that a percept may arise precisely at this macroscopic level.
6.3 Symmetry between possible states
Let us imagine that the sensory input supports two possible interpretations:
M_A
M_B.
Their initial dynamic stability may be approximately the same:
stability(M_A) ≈ stability(M_B).
The system finds itself in a situation where it is not unambiguously determined which state will be realized.
This situation can be understood as a dynamic symmetry.
It does not necessarily mean an exact mathematical symmetry of all neurons.
It means that several macroscopic possibilities have a similar dynamic accessibility.
6.4 Spontaneous symmetry breaking
If the state is perfectly balanced, a small local fluctuation may create an initial difference:
activity_A =
activity_B + ε.
This deviation may arise, for example, from:
stochastic spike timing,
phase difference,
synaptic variability,
previous state,
recurrent fluctuation.
If the network contains amplifying feedback:
ε
->
local advantage
->
recurrent amplification
->
stronger suppression of competitor
->
further advantage.
The result may be:
M_A >> M_B.
An originally approximately symmetric state ends up in one particular configuration.
This is a working neuronal analogy of spontaneous symmetry breaking.
6.5 Symmetric rules, asymmetric outcome
An important principle is:
symmetric local rules
do not imply
symmetric global outcome.
A network may have the same parameters for two competing populations:
W_A = W_B,
threshold_A = threshold_B,
input_A ≈ input_B.
Nevertheless a particular run ends, for example, in the state:
M_A.
Another run may end in:
M_B.
The result therefore need not be written into the architecture in advance.
It may arise from the dynamics of the system.
6.6 Stochasticity as an initiating mechanism
Spontaneous stochasticity here acquires a concrete function.
It need not create the structure of the percept itself.
It may only determine which of the already available dynamic possibilities gains the initial advantage.
Schematically:
structured state space
+
small stochastic fluctuation
->
state selection.
That is:
stochasticity != percept content
but:
stochasticity
can influence
percept selection.
This allows the system to decide even in a situation where the sensory evidence is not sufficient for an unambiguous deterministic choice.
6.7 Oscillatory phase as a source of asymmetry
The initial asymmetry need not be created by noise alone.
If two populations exist in a different phase configuration:
φ_A != φ_B,
then the same input may have a different effect.
For example:
excitability_A > excitability_B
at a particular moment.
This gives rise to:
same sensory evidence
+
different internal phase
->
different initial advantage.
Recurrence may subsequently amplify this difference.
The oscillatory state may therefore influence which percept will be realized.
6.8 Non-commutativity as a source of asymmetry
History may create an asymmetry in a similar way.
The sequence:
X -> Y
may prepare the system into the state:
S_XY
while:
Y -> X
leads into:
S_YX.
If:
S_XY != S_YX,
then the same ambivalent input Z may end up as:
S_XY + Z -> M_A
while:
S_YX + Z -> M_B.
The selection of the percept therefore depends on the path the system has traversed.
6.9 Recurrent amplification
A small asymmetry by itself need not be sufficient.
For the formation of a coherent macrostate, positive feedback is important.
For example:
population A
->
recurrent excitation of A
and at the same time:
population A
->
inhibition of B.
Then a small difference:
A = 0.51
B = 0.49
may be gradually amplified:
0.51 / 0.49
->
0.60 / 0.40
->
0.75 / 0.25
->
0.90 / 0.10.
The result is a coherent state.
6.10 Competition and winner-take-most
DPSH need not require an absolute winner-take-all.
In a biological system it may be more realistic to have:
winner-take-most.
The dominant state:
M_A
may suppress the alternative representation:
M_B
without removing it completely.
This allows:
residual alternatives,
later switching,
ambiguity,
perceptual reversals.
Such a system is more suitable for metastability than an absolutely stable winner-take-all.
6.11 Inhibition as an organizing mechanism
Inhibition need not be merely a mechanism reducing activity.
It may create structure by constraining the possible simultaneous configurations.
For example:
A inhibits B
B inhibits A.
A dynamic competition arises.
In a more complex network:
local excitation
+
lateral inhibition
may create selective assemblies.
DPSH therefore regards inhibition as one of the conditions of the self-organization of a coherent state.
6.12 Constraint satisfaction
Self-organization can also be interpreted as the dynamic solution of a system of constraints.
Every part of perception may create a local constraint:
color,
shape,
motion,
position,
memory context,
prediction.
A possible global state must be compatible with as many of these constraints as possible.
The network need not explicitly compute:
optimize(global_objective).
Instead, local interactions may gradually destabilize inconsistent configurations and stabilize compatible ones.
Schematically:
many local constraints
->
recurrent interaction
->
incompatible states decay
->
compatible macrostate survives.
6.13 Predictive processing as a selection pressure
Predictive processing may markedly influence this dynamics.
Let us imagine two candidate states:
M_A
M_B.
Both explain part of the input.
But their prediction error is different:
ε_A < ε_B.
Then it may hold that:
stability(M_A) > stability(M_B).
The predictive mechanism therefore need not create the percept.
It may change the "dynamic landscape" so that some states are more stable than others.
6.14 The dynamic landscape
For an intuitive description we may introduce the metaphor of a dynamic landscape.
The state of the system:
S(t)
moves in a space of possibilities.
Some regions are:
unstable,
transient,
metastable,
highly stable.
Sensory input, experience, oscillatory phase and plasticity may change this landscape.
Schematically:
state-space geometry =
F(
connectivity,
history,
prediction,
oscillatory state,
current input
).
Perception is then not merely the selection of a label.
It is the movement of a dynamic system through a landscape of possible interpretations.
6.15 Symmetry breaking and the attractor landscape
If two similarly deep dynamic regions exist:
basin A
basin B,
then a small perturbation may determine which of them the system enters.
After entry:
recurrent dynamics
maintains the state for a certain time.
This provides a mechanism:
ambiguity
->
fluctuation
->
basin selection
->
perceptual stabilization.
6.16 Why metastability is preferable to absolute stability
An absolutely stable attractor might be unsuitable for living perception.
If the system once enters:
M_A
and cannot easily leave it, it will not be able to respond to a change in the world.
DPSH therefore expects:
sufficient stability
+
possibility of transition.
This is metastability.
Self-organization must therefore create a state that is:
coherent enough to persist
but:
flexible enough to change.
6.17 Symmetry breaking and perceptual reversals
Ambivalent stimuli provide a natural test.
The same external input may lead to:
M_A
and later to:
M_B.
If the sensory input has not changed, the change must originate from the internal dynamics.
A candidate mechanism:
adaptation
+
phase drift
+
stochastic fluctuation
->
destabilization of M_A
->
symmetry restored temporarily
->
selection of M_B.
Such dynamics is compatible with perceptual switching.
6.18 Order parameter
In physics, spontaneous symmetry breaking is often described by means of a macroscopic order parameter.
For a neuronal network we may analogously define:
Ψ(S).
For two competing populations, for example:
Ψ =
(A - B) / (A + B).
Then:
Ψ ≈ 0
means a balanced state.
Conversely:
Ψ >> 0
means the dominance of A
and:
Ψ << 0
the dominance of B.
Symmetry breaking can thereby be measured directly.
6.19 Multidimensional order parameter
For a more complex percept a single scalar is not sufficient.
We may define a vector:
Ψ =
(
coherence,
phase_structure,
cluster_occupancy,
prediction_consistency,
state_separability
).
Such a parameter may describe the macroscopic organization without the need to track every individual neuron.
6.20 Phase transition
Symmetry breaking may be associated with a phase transition.
With the gradual increase of some parameter:
λ
the network may remain for a long time in a disordered regime.
Then in the vicinity of:
λ_c
a qualitative change occurs:
distributed weak activity
->
coherent macrostate.
Candidate parameters:
recurrent gain,
sensory evidence,
coupling strength,
oscillatory coherence,
inhibition strength.
DPSH examines the possibility that the formation of a perceptual state may have precisely the character of such a dynamic transition.
6.21 Relation to ignition
The Global Neuronal Workspace uses the notion of ignition for the abrupt amplification and global availability of a representation.
DPSH proposes a possible distinction:
local symmetry breaking
->
coherent metastable percept
->
workspace ignition.
It is, however, also possible that some aspects of ignition constitute directly a macroscopic phase transition of a broader dynamics.
This must be tested, not assumed.
6.22 Self-organization without a central controller
This is a fundamental architectural requirement.
Cognia must not resolve the percept by, for example:
controller.select(best_state).
Such a mechanism would merely move the problem one level higher.
Instead, the selection must arise from:
local excitation,
inhibition,
recurrence,
timing,
stochasticity,
local plasticity.
A controller may later modulate the dynamics.
It should not, however, explicitly construct the percept.
6.23 Local rules and global order
Every unit uses simple local rules:
if spike arrives:
change local state
if threshold/hazard condition:
generate spike
if pre/post timing:
modify synapse.
Nevertheless there may arise a global:
assembly,
attractor,
phase relation,
metastable percept.
This is one of the main principles of DPSH:
The complexity of the percept need not be explicitly encoded in the complexity of the individual neuron.
6.24 Symmetry breaking and the number of neurons
An interesting but so far speculative question arises here.
A large population of simple autonomous units may create richer statistics of fluctuations and more possible collective configurations than a small number of very complex units.
It is possible that:
many simple units
->
richer emergent macrostate space.
This is not, however, a present conclusion of DPSH.
It is a separate scaling hypothesis that can be tested experimentally later.
6.25 Symmetry breaking and experience
Learning changes the dynamic landscape.
Before experience:
basin_A ≈ basin_B.
After repeated encounters with A:
plasticity
->
basin_A deepens.
Then the same ambivalent input more often ends in:
M_A.
Previous experience thereby manifests itself as a bias of future symmetry breaking.
6.26 Intuition as rapid macrostate selection
This mechanism provides a possible interpretation of intuition.
After long-term learning, the dynamic landscape may contain stable or easily accessible regions.
A new complex input may move the network very rapidly:
S0 -> M_A
without an explicit symbolic reasoning process.
The resulting state may directly influence:
action.
The system may thus "know" which macrostate corresponds to the situation, without having a globally available explicit causal reconstruction of the process by which it arrived at it.
This is a working functional interpretation of intuitive decision-making.
6.27 Symmetry breaking and the Perceptual Manifold
In the Perceptual Manifold the possible percepts can be understood as regions:
M_A,
M_B,
M_C.
Symmetry breaking is then the process:
ambiguous region
->
trajectory divergence
->
one selected basin.
The Perceptual Manifold is therefore not merely a map of finished percepts.
It also contains the boundaries and transitional regions where the dynamics may decide between alternative interpretations.
6.28 Basin boundaries
The regions close to the boundaries between states will be especially important.
If:
S(t) near boundary(M_A, M_B),
a small perturbation may lead to a different outcome.
Sensitivity to perturbation can be measured as:
sensitivity(S) =
P(different final macrostate | small perturbation).
DPSH expects high sensitivity precisely in the transitional regions.
6.29 Experiment SB1 – perfectly balanced competition
We create two symmetric populations:
A
B
with:
same size,
same weights,
same input,
same thresholds.
The system receives an ambivalent input.
We track:
whether symmetry breaks,
time to selection,
final state,
repeatability.
With stochasticity we expect the distribution:
P(M_A) ≈ P(M_B)
under perfectly symmetric conditions.
6.30 Experiment SB2 – small bias
To the previous experiment we add:
ε.
For example:
input_A = input + ε.
We measure:
P(M_A | ε).
If the system amplifies small differences, a sigmoidal or otherwise nonlinear dependence should arise:
ε
->
selection probability.
This will make it possible to measure the sensitivity of dynamic decision-making.
6.31 Experiment SB3 – stochasticity sweep
We will vary:
σ.
For each value we measure:
symmetry breaking time,
stability,
switching rate,
accuracy under weak evidence.
We may perhaps expect:
σ too low
->
slow/no selection
σ moderate
->
flexible selection
σ high
->
unstable selection.
This directly connects chapter 3 with symmetry breaking.
6.32 Experiment SB4 – phase bias
We keep the sensory input perfectly symmetric.
We change only the initial relative phase:
Δφ.
We track:
P(M_A | Δφ).
If the phase configuration systematically biases the result, oscillations constitute a causal component of macrostate selection.
6.33 Experiment SB5 – history bias
Before the ambivalent input we create two different histories:
history_A
history_B.
We then present an identical stimulus:
X.
We measure:
P(M_A | history_A)
versus:
P(M_A | history_B).
This tests the connection:
noncommutativity
->
state preparation
->
symmetry breaking.
6.34 Experiment SB6 – recurrent gain
We will vary the strength of the recurrent excitation:
g_rec.
At a low value there may be:
no stable selection.
At an intermediate value:
metastable percept.
At too high a value:
rigid attractor / pathological locking.
We measure:
order parameter Ψ,
state lifetime,
transition probability,
perturbation recovery.
In this way a phase transition can be sought.
6.35 Experiment SB7 – inhibition strength
Similarly:
g_inh.
Too weak an inhibition may cause:
simultaneous activation.
Too strong:
global suppression.
An intermediate regime may allow:
selective coherent assemblies.
This will determine whether competition genuinely contributes to the formation of an unambiguous macrostate.
6.36 Experiment SB8 – perceptual reversal
The network receives a constant ambivalent stimulus over a longer period.
We track whether the following arises spontaneously:
M_A
->
M_B
->
M_A
-> ...
If so, we analyse:
adaptation,
stochastic fluctuations,
phase drift,
transition timing.
Such an experiment is especially suitable for the study of metastability.
6.37 Experiment SB9 – perturbation of the macrostate
After the stabilization of:
M_A
we carry out a small local perturbation.
For example:
activate subset supporting B
or:
inhibit subset supporting A.
We measure the minimal perturbation needed for:
M_A -> M_B.
In this way the depth and robustness of the dynamic basin can be quantified.
6.38 Experiment SB10 – remove central selection
If the experimental architecture contains a controller, we create a variant in which the controller has no access to the choice of the percept.
We compare:
local self-organized selection
against:
centrally selected state.
The aim is to show whether the network is able to create a coherent state without an explicit global selection operation.
6.39 Falsification criteria
The strong hypothesis of self-organization and symmetry breaking will be weakened if:
- a coherent macrostate does not arise without an explicit central selector,
- small local perturbations are not able to bias the result,
- recurrence does not amplify the initial differences,
- inhibition/competition is not relevant for the selection,
- the initial phase or history has no influence on an ambivalent percept,
- the resulting state is simply a linear function of the input,
- no measurable order parameter exhibits a transition between a disordered and an organized state,
- metastable macrostates do not exhibit a robust basin-like structure.
In such a case symmetry breaking would not be a suitable explanation of the formation of the percept and would have to be replaced by a simpler mechanism.
6.40 Chapter research hypothesis
We formulate the partial hypothesis H5:
H5 – Self-Organized Symmetry Breaking Hypothesis
In a recurrent globally clockless neuronal network, competition among several dynamically available interpretative states may, through local stochasticity, temporal asymmetry, recurrent amplification and inhibition, lead to spontaneous symmetry breaking and to the formation of a single coherent metastable macrostate.
The stronger prediction reads:
If perceptual selection is genuinely self-organized, then with an ambivalent input it must be possible to change systematically the probability of the resulting percept by means of small local perturbations, a change of history or of relative phase, without there being a central mechanism explicitly selecting the resulting state.
The hypothesis does not claim:
symmetry breaking = consciousness.
It claims:
competing possibilities
+
local interactions
->
spontaneous macrostate selection.
It thereby provides a candidate mechanism for how a single coherent perceptual state may arise out of distributed activity without a central constructor.
7. The metastable Perceptual Manifold
7.1 From the selection of a state to its existence
The preceding chapters described mechanisms that may enable the formation of a coherent macrostate:
- spontaneous stochasticity,
- temporal organization by oscillations,
- non-commutative dynamics,
- recurrent amplification,
- spontaneous symmetry breaking.
It is now necessary to define what exactly it means that the network has created a perceptual state.
The Dynamic Perceptual State Hypothesis does not assume that a percept corresponds to:
a single neuron,
nor to:
a single layer,
nor to:
a single static pattern of activation.
The working hypothesis is stronger:
A percept corresponds to a distributed dynamic macrostate that persists for a certain time despite microscopic changes of individual neuronal activities and that influences the further evolution of the system.
7.2 The global state of the system
Let us consider a network of N dynamic units.
Its instantaneous global state can be written:
S(t) =
(
s1(t),
s2(t),
...,
sN(t)
).
Every local state si(t) may contain, for example:
membrane potential,
refractory state,
adaptation,
spike history,
local phase context,
modulatory state.
If we also include the synaptic and oscillatory dynamics, a more complete state may be:
S(t) =
{
neural states,
synaptic states,
oscillatory states,
modulatory states
}.
This global state is not a centrally represented data structure.
It is an analytical description of the entire physical system.
7.3 The state space
The set of all states the system can attain forms the state space:
Ω.
During its existence the network creates a trajectory:
S(t0)
->
S(t1)
->
S(t2)
->
...
->
S(tn).
In a continuous description:
S : t -> Ω.
The object of research in DPSH is therefore not merely:
output(t),
but above all:
trajectory through Ω.
7.4 A percept is not a point
A simple representation might correspond to a point:
S_A.
Such a model is, however, too rigid.
The same percept may be realized by many microscopically different configurations:
S_A1,
S_A2,
S_A3,
...
We therefore define a perceptual region:
M_A ⊂ Ω.
If:
S(t) in M_A,
we say that the system is in the dynamic state corresponding to percept
A.
The individual neurons may meanwhile be changing continuously.
7.5 Macroscopic identity
Two microstates:
S1
S2
may be very different at the level of individual spikes.
Nevertheless they may belong to the same macrostate:
S1 in M_A
S2 in M_A.
The identity of the percept therefore need not require:
identical spikes.
It requires the preservation of certain macroscopic relations.
For example:
similar population geometry,
similar phase organization,
similar transition tendencies,
similar downstream effect,
similar predictive content.
This is one of the key principles of DPSH:
Perceptual identity may be invariant with respect to part of the microscopic neuronal variability.
7.6 What metastability means
The state M_A is not necessarily a permanent attractor.
It is metastable.
This means:
S(t) in M_A
for the duration:
t0 < t < t1,
but there is a non-zero probability:
P(M_A -> M_B) > 0.
The state therefore satisfies two properties at once:
persistence
and:
transition capability.
This is important for perception.
A state that is too unstable would not be able to maintain a coherent percept.
A state that is too stable would not be able to respond to a change in the world.
7.7 Metastability versus a fixed-point attractor
A fixed-point attractor:
S(t) -> S*
and then:
S(t + dt) ≈ S*.
A dynamic metastable state may, by contrast, exhibit:
S1 -> S2 -> S3 -> S4 -> ...
where:
all Si in M_A.
A percept may therefore be dynamic inside its own region.
This is important.
DPSH does not assume that the stability of a percept means the cessation of dynamics.
7.8 Metastable trajectories
Some percepts may not correspond even to a static region.
They may be characterized by a typical trajectory:
T_A.
For example:
S1 -> S2 -> S3 -> S4
may represent one stable dynamic cycle or sequence.
Another realization of the same percept:
S1' -> S2' -> S3' -> S4'
may be different at the microscopic level, but geometrically similar.
A percept can therefore be described not only as:
region M_A
but also as:
family of trajectories T_A.
7.9 Dynamic invariance
DPSH assumes that a percept must have a certain degree of invariance.
For example, a change of a few spikes:
perturbation ε
is not to create a different percept immediately.
If:
S(t) + ε
remains in:
M_A,
then the representation is robust.
This property can be measured:
robustness(M_A).
A stronger perturbation may cause:
M_A -> M_B.
We thereby obtain an experimentally measurable boundary of the perceptual region.
7.10 Basin of attraction
For a perceptual state we may define a basin:
B_A.
It is the set of initial states that have a high probability of transition into:
M_A.
That is:
S0 in B_A
->
P(S(t) -> M_A) high.
Learning may change:
size(B_A),
shape(B_A),
depth(B_A).
Familiar percepts may have larger or more easily accessible basins than unfamiliar configurations.
7.11 Dynamic geometry
The Perceptual Manifold is not a fixed geometry.
Its structure may be a function of:
W(t),
phase(t),
sensory input,
internal context,
neuromodulation.
That is:
M = M(t).
The same network may have a different dynamic geometry in a different context.
7.12 Learning as a deformation of the manifold
An experience:
E
causes plasticity:
W -> W'.
This changes the dynamics:
F -> F'.
And hence also the state space:
M_before
->
M_after.
Learning can therefore be understood as:
a deformation of the geometry of the available internal states and of the probabilities of the transitions between them.
This interpretation is fundamental for DPSH.
The network does not learn only:
input -> output.
It learns:
which internal states are easy to reach,
which states are stable,
how transitions occur.
7.13 Deep State Learning
The working term Deep State Learning designates the hypothesis that plasticity may change the structure of the internal dynamic space even outside a direct supervised mapping of input onto output.
Schematically:
experience
->
state trajectory
->
local plasticity
->
altered transition geometry.
A stronger version assumes:
spontaneous internal dynamics
->
plasticity
->
continued restructuring.
This part must be experimentally separated from ordinary learning during external stimulation.
7.14 Perceptual Manifold
We define a working object:
P = Perceptual Manifold.
P is not a single representation.
It is a structure:
P = {
perceptual regions,
trajectories,
transition probabilities,
phase relations,
learned constraints,
basin geometry
}.
It therefore contains not only:
"what the system is currently perceiving",
but also:
"into which states it may pass"
and
"how easy these transitions are".
7.15 The local percept and the global manifold
An individual percept:
M_A
is only a part of:
P.
For example:
M_face
M_table
M_hand
M_space
M_motion.
These states need not exist in isolation.
They may be mutually interconnected.
Global perceptual experience may be the result of their simultaneous dynamic configuration.
7.16 The integrated state
DPSH assumes the possibility that a percept is not a simple sum:
M_A + M_B + M_C.
The interactions may create a new state:
M_ABC
which is not reducible to independent components.
That is:
F(A, B)
!=
F(A) + F(B).
This is a further consequence of nonlinearity and recurrence.
7.17 Binding without a central binder
If different properties:
color,
shape,
position,
motion
are dynamically compatible and enter into a common metastable configuration, there need not exist a separate module:
bind_features().
Binding may be an emergent property of the common macrostate.
DPSH does not, however, claim that oscillatory synchronization is itself a universal mechanism of binding.
The previous literature review showed that the simple binding-by-synchrony claim has significant counterarguments.
We therefore work with a more general principle:
coordinated dynamics
->
integrated state.
7.18 The perceptual state as context
As soon as there arises:
M_A,
it becomes part of the input conditions for further processing.
A new stimulus X is therefore not interpreted as:
Y = F(X),
but as:
Y = F(X, M_A).
If the previous state was:
M_B,
it may hold that:
F(X, M_A)
!=
F(X, M_B).
This provides a functional definition of an internal percept:
A state is perceptually relevant if its existence changes the interpretation of a subsequent input.
7.19 Persistence without an explicit memory cell
A perceptual context may persist:
stimulus
->
M_A
->
blank interval
->
M_A-like dynamics.
It is not necessary that there exists a cell:
memory_A = 1.
Information may remain distributed in:
population state,
phase relations,
recurrent activity,
short-term synaptic state,
dynamic trajectory.
This is one of the key differences between explicit memory and a dynamic perceptual state.
7.20 Active and activity-silent components
DPSH need not assume that the whole percept must always be represented by strong ongoing firing.
Part of the context may exist in:
synaptic state,
altered excitability,
phase configuration,
latent connectivity.
Active dynamics may reactivate some parts of this state.
The Perceptual Manifold may therefore contain:
active components
as well as:
latent components.
7.21 The percept as a predictive state
If the state:
M_A
represents an internal model of a certain situation, it should generate expectations.
That is:
M_A
->
prediction P_A.
The following sensory input may be:
compatible
or:
incompatible.
A compatible input supports:
persistence(M_A).
An incompatible input increases the probability of:
transition(M_A -> M_B).
The Perceptual Manifold is thereby connected with predictive processing.
7.22 The percept as a compression of history
The present state:
M_A
may be the result of a long previous sequence:
X1 -> X2 -> X3 -> ... -> Xn.
It need not, however, explicitly retain all the individual events.
It may represent a compressed consequence of their history.
That is:
history
->
manifold state.
This may be computationally significant.
Other modules need not analyse the whole past anew.
They may respond to the present dynamic state.
7.23 Intuition as a reading of the manifold
This mechanism provides a further possible interpretation of intuition.
Complex experience deforms the Perceptual Manifold over the long term.
A new situation may then very rapidly move the system:
S0 -> M_warning.
An action or evaluation module may respond:
M_warning -> avoid
without it being necessary to reconstruct explicitly:
X1 -> X2 -> ... -> reason.
Intuition might thus be a function of the rapid reading of the learned geometry of the internal state space.
7.24 The state as a shared internal reality
If several modules respond to the same dynamic structure:
P
a shared internal frame of reference may arise.
For example:
perception,
memory,
value,
action,
language
may obtain different projections:
O_i = f_i(P).
One module derives from the dynamic state:
"object is reachable",
another:
"object is dangerous",
a further one:
"object is named chair".
None of them need reconstruct the whole sensory input anew.
7.25 Projections of the manifold
Formally, a module i can be understood as a transformation:
O_i = G_i(P).
Different modules have different observables.
The Perceptual Manifold may therefore function as a:
common latent dynamic state
with multiple functional projections.
7.26 Percept formation versus global accessibility
It is necessary to separate:
existence M_A
from:
access(M_A).
A perceptual state may exist locally or in a distributed manner without being automatically available to all subsystems.
The Global Workspace may subsequently enable:
M_A
->
global broadcast.
DPSH therefore distinguishes:
percept formation
and:
global access.
7.27 A percept without report
This distinction allows the experimental possibility:
perceptual dynamics present
but:
explicit report absent.
The network may, for example:
use state for action
without making it available to the:
language/report module.
Implicit or intuitive processing can be modelled in this way.
7.28 Causal relevance
Mere decodability is not sufficient.
If:
A
can be decoded from the activity, this does not automatically mean that
the system uses the state A.
DPSH therefore requires a causal test.
If we perturb:
M_A -> M_B,
there must be a change in:
future interpretation
or
behavior.
Only then can we claim that the state is functionally relevant.
7.29 Operational definition of a perceptual state
For experimental purposes we will designate a state M_A as perceptual
only if it satisfies at least the following criteria.
1. Decodability
From the population dynamics it is possible to distinguish:
M_A
from:
M_B.
2. Persistence
The state persists longer than the immediate sensory event.
3. Robustness
Small perturbations do not destroy it immediately.
4. Metastability
It is not a rigid permanent fixed point.
5. History dependence
Its formation or interpretation depends on the previous state.
6. Causal relevance
A perturbation of the state changes further processing or decisions.
7. Sensory grounding
The state is systematically related to sensory experience or context.
8. Generalization
It is not the mere memorization of one particular input pattern.
7.30 The separability metric
For two states:
M_A
M_B
their separability can be measured:
D(M_A, M_B).
For example by means of:
centroid distance,
classifier accuracy,
manifold distance,
trajectory distance.
If:
D >> within-state variance,
the states are distinguishable.
7.31 Dwell time
For a metastable state we define:
T_dwell(M_A).
It is the time for which the system remains in the given region before a transition.
The distribution:
P(T_dwell)
may characterize the dynamic stability of the percept.
7.32 Transition matrix
For a set of states:
{M1, M2, ..., Mk}
we may define:
P_ij =
P(M_i -> M_j).
A transition matrix arises:
P.
It characterizes the dynamic geometry of the Perceptual Manifold.
7.33 Transition entropy
The degree of predictability of transitions can be described by an entropy:
H_transition.
A low value:
rigid dynamics.
A very high one:
chaotic/unstructured transitions.
DPSH assumes the possibility of an intermediate regime:
structured but flexible dynamics.
7.34 Trajectory reproducibility
Upon repeated presentation of the same context, the same spikes need not arise.
Nevertheless a similar macroscopic trajectory may exist.
We measure:
similarity(T_A1, T_A2).
This will make it possible to test whether the percept exists at a higher level than the particular spike sequence.
7.35 Perturbation stability
After the formation of:
M_A
we apply a perturbation:
ε.
We track:
M_A
->
recovery to M_A
or:
M_A
->
transition to M_B.
We thus obtain the boundary of dynamic stability.
7.36 Experiment M1 – A / blank / ambiguous X
The basic perceptual experiment:
A
->
blank
->
X_ambiguous
and:
B
->
blank
->
X_ambiguous.
X_ambiguous is identical in both cases.
The correct response must depend on the previous state:
A -> X -> Y_A
B -> X -> Y_B.
During the blank interval we measure:
S(t).
We look for two distinguishable regions:
M_A
M_B.
If the state during the blank predicts the subsequent decision, we have functional evidence of an internal perceptual context.
7.37 Experiment M2 – state perturbation
After the formation of:
M_A
we carry out a perturbation towards:
M_B.
If the subsequent response to:
X_ambiguous
changes from:
Y_A
to:
Y_B,
the state is causally relevant.
7.38 Experiment M3 – microscopic variability
We repeat the same stimulus many times with stochasticity.
We compare:
spike-level variability
and:
macrostate similarity.
DPSH assumes that:
high microscopic variability
may coexist with:
high macrostate stability.
7.39 Experiment M4 – remove recurrence
We compare:
recurrent network
and:
recurrence reduced/removed.
If recurrence plays a fundamental role in metastability, we expect a decline in:
persistence,
robustness,
context retention.
7.40 Experiment M5 – explicit memory control
We create a control model:
memory_cell = previous_context.
This may solve the task:
A -> blank -> X
explicitly.
We compare it with the dynamic network.
The aim is not to prove that explicit memory is worse.
The aim is to find out whether the dynamic network exhibits additional properties:
spontaneous transitions,
generalization,
phase sensitivity,
metastability,
perturbation recovery.
7.41 Experiment M6 – the manifold after learning
We measure the state space:
before learning
and:
after learning.
We track:
number of clusters,
basin geometry,
transition probabilities,
spontaneous/evoked similarity.
If experience genuinely deforms the manifold, the change must be quantifiable.
7.42 Experiment M7 – spontaneous replay
After learning we remove the input:
I_external = 0.
We track whether the spontaneous trajectories visit regions:
M_A,
M_B,
...
that were previously associated with sensory experiences.
This connects the metastable manifold with the hypothesis of spontaneous learning.
7.43 Experiment M8 – context switching
We place the system into different contexts:
C1
C2.
The same stimulus:
X
may lead to:
C1 + X -> M_A
and:
C2 + X -> M_B.
If so, the manifold is not a fixed lookup table of inputs.
It is contextually dynamic.
7.44 Experiment M9 – generalization
We train the network on a set of variants of a percept:
A1,
A2,
A3.
We then present a new one:
A4.
If:
A4 -> M_A,
then the perceptual region captures a more general structure than the exact memorized pattern.
7.45 Experiment M10 – manifold fragmentation
We will progressively disrupt:
phase,
timing,
recurrence,
stochasticity.
We track whether:
M_A
remains one coherent region or breaks up into fragments.
In this way it is possible to determine which mechanisms hold the perceptual structure together.
7.46 The Perceptual Manifold and scaling
With a growing number of neurons:
N
the number of available dynamic configurations may grow.
It is not, however, automatically true that:
larger N -> better percept.
It is necessary to measure:
effective dimensionality,
redundancy,
manifold separability,
computational cost.
There may perhaps be an optimal relation among:
number of units,
connectivity,
stochasticity,
temporal structure.
7.47 The phenomenal question
If the system creates a:
robust,
integrated,
history-dependent,
predictive,
causally effective
metastable Perceptual Manifold,
it is still not possible to derive from this directly:
phenomenal experience.
DPSH therefore preserves the boundary:
functional perceptual state
!= proven quale.
The stronger phenomenal hypothesis remains open.
7.48 Why this chapter is central
The preceding mechanisms:
stochasticity,
oscillations,
noncommutativity,
symmetry breaking
are means.
The metastable Perceptual Manifold is the object whose formation we are attempting to explain.
If such a dynamic structure does not arise in experiment, there is no point in proceeding to the Global Workspace or to the phenomenal interpretation.
This chapter is therefore the experimental centre of the whole of DPSH.
7.49 Falsification criteria
The hypothesis of the metastable Perceptual Manifold will be weakened if:
- internal states cannot be reliably separated,
- the state does not persist across a blank interval,
- a perturbation of the internal state does not change the subsequent decision,
- the same percept requires an almost identical microscopic activity,
- the state-space structure does not change after learning,
- the previous state does not change the interpretation of an identical subsequent input,
- an explicit simple memory variable explains all the observed effects without a loss of the relevant properties,
- the dynamic state carries no information beyond the current input,
- no region of the state space exhibits a metastable character.
In such a case the central conception of DPSH must be fundamentally reformulated.
7.50 Chapter research hypothesis
We formulate the partial hypothesis H6:
H6 – Metastable Perceptual Manifold Hypothesis
Perceptually relevant information may be represented as a distributed metastable region or family of trajectories of the global state space of a recurrent neuronal network. The identity of such a state persists across the microscopic variability of individual spikes, preserves part of the sensory history and causally influences the interpretation of subsequent inputs.
The stronger prediction:
If a percept genuinely corresponds to a metastable dynamic structure, then it must be possible to decode the internal context from the population trajectory, this state must persist without immediate sensory input, and its targeted perturbation must change the subsequent interpretation or behaviour of the system.
At this stage DPSH therefore claims:
percept
=
functionally relevant metastable population dynamics
not:
percept
=
static activation pattern.
And it still does not claim:
metastable state
=
phenomenal quale.
This last relation remains a separate open hypothesis.
8. Hysteresis and the continuity of the inner world
8.1 The internal state does not start over
The Dynamic Perceptual State Hypothesis assumes that a perceptual system does not create a new internal world from the beginning at every change of sensory input.
Instead, the new input acts upon an already existing state:
S(t + dt) = F(S(t), I(t)).
This means that:
current perception
is not merely a function of:
current sensory input,
but also of:
previous internal state.
This principle creates continuity.
The system does not pass through:
no world
->
new world
->
no world
->
new world.
An internal dynamic representation exists continuously and changes in time.
8.2 Continuity versus frame-based perception
A simple discrete model of perception may be represented as:
frame_1 -> representation_1
frame_2 -> representation_2
frame_3 -> representation_3.
Such a model may create accurate representations of individual inputs, but by itself it does not explain why these representations are part of one continuous internal world.
DPSH prefers the description:
S0
->
S1
->
S2
->
S3
where every new state arises by a transformation of the previous one.
Sensory input is one of the factors of this transformation.
It is not the sole source of the state.
8.3 Hysteresis
Hysteresis means that the response of the system depends on its previous trajectory.
Let us imagine a parameter:
x
which gradually increases.
The system passes:
M_A -> M_B
at:
x = θ_AB.
We then begin to decrease x.
The transition:
M_B -> M_A
need not occur at the same value.
It may hold that:
θ_AB != θ_BA.
This means that the current value of the input alone is not sufficient to determine the state.
We must know the:
history.
8.4 Hysteresis as macroscopic memory
Hysteresis constitutes a form of memory that need not be realized by an explicit memory cell.
Information about the past is contained in the present dynamic configuration.
That is:
history
->
current macrostate.
The system need not retain:
previous_value = X.
It suffices that the previous development changed the:
population state,
synaptic state,
phase relations,
excitability,
attractor occupancy.
The past thereby becomes a physical property of the present.
8.5 Perceptual continuity
One of the consequences of hysteresis may be the stability of a percept during a short-term loss of sensory evidence.
An object may, for example, be temporarily occluded.
The sensory input associated with the object drops abruptly:
I_object -> 0.
Nevertheless it need not immediately hold that:
percept_object -> 0.
If the system remains in the region:
M_object,
then the object may remain part of the internal representation even during a short loss of input.
8.6 Persistence is not the same as immutability
The continuity of the inner world does not mean that the representation must be static.
It may hold that:
M_object(t0)
!=
M_object(t1)
at the microscopic level,
but both states may still represent the same object in a dynamic context.
A percept may be continuously updated without its identity being extinguished.
8.7 Object permanence as a dynamic phenomenon
DPSH allows object permanence to be interpreted as a property of the dynamic state.
Before occlusion:
visible object
->
M_object.
After occlusion:
sensory evidence decreases,
but:
prediction + previous state
->
persistence of M_object.
The object therefore does not necessarily remain "in memory" as an explicit symbol.
It may remain as an expected component of the dynamic model.
8.8 Prediction as a mechanism of continuity
If the internal state represents an object, it may generate an expectation:
object continues to exist.
The sensory system may then expect, for example:
predicted position,
predicted motion,
predicted reappearance.
If sensory evidence is briefly absent, the prediction error need not be sufficient for the immediate destabilization of the state.
This gives rise to:
perceptual persistence.
8.9 Breakdown of the percept under long-term conflict
Continuity cannot, however, be unlimited.
If sensory evidence contradicts the internal state over the long term:
prediction_error >> threshold,
then:
stability(M_A) decrease.
must follow.
Ultimately:
M_A -> M_B
or:
M_A -> unresolved state.
This prevents the inner world from becoming completely severed from reality.
8.10 Hysteresis versus hallucination
This distinction is fundamental for DPSH.
Too low a persistence:
world representation collapses with every missing input.
Too high a persistence:
internal state ignores contradictory evidence.
The first regime leads to an unstable percept.
The second may lead to self-sustaining internal states that are no longer sufficiently anchored in the environment.
The hypothesis therefore expects an optimal region between:
excessive instability
and:
excessive persistence.
8.11 State inertia
We may introduce a property:
inertia(M_A).
It expresses how strong a change of input is needed to leave the state.
Low inertia:
small perturbation -> transition.
High inertia:
large perturbation required.
Perceptual continuity requires a non-zero but not infinite inertia.
8.12 The hysteresis loop
Experimentally, hysteresis can be measured by changing the input upwards and downwards.
For example:
ambiguous stimulus parameter = x.
With:
x increasing
we measure the transition:
A -> B.
With:
x decreasing
we measure:
B -> A.
If:
θ_AB != θ_BA,
a hysteresis loop arises.
Its width:
H = |θ_AB - θ_BA|
may be a simple metric of state dependence.
8.13 Hysteresis and non-commutativity
Hysteresis is closely connected with non-commutative dynamics.
The sequence:
A -> X
need not be equivalent to:
B -> X.
The same final input X acts upon two different previous states.
Therefore:
F(M_A, X)
!=
F(M_B, X).
Hysteresis is thus a macroscopic consequence of the fact that history has changed the current state of the system.
8.14 Hysteresis and symmetry breaking
After the spontaneous selection of:
M_A
the system may remain in this state even after the original small
advantage of A has disappeared.
This is fundamental.
Symmetry breaking creates the state.
Hysteresis may stabilize it for a certain time.
Schematically:
ambiguity
->
symmetry breaking
->
M_A
->
hysteretic persistence.
8.15 Hysteresis and metastability
Hysteresis must not lead to permanent locking.
DPSH therefore associates it with metastability.
It holds that:
previous state biases future state,
but:
sufficient new evidence
->
transition.
This creates a continuous yet adaptive inner world.
8.16 A percept is not a reconstruction of every moment
This part of the hypothesis leads to an important consequence.
The brain need not resolve at every sensory moment:
"What does the whole world look like right now?"
Instead it may resolve:
"What has changed relative to what I already assume?"
That is:
existing internal world
+
sensory update
->
modified internal world.
This is markedly more efficient than a complete reconstruction from scratch.
8.17 The Perceptual Manifold as a persistent structure
The Perceptual Manifold is not created anew at every input.
It is continuously deformed.
Formally:
P(t + dt)
=
G(P(t), I(t), prediction, plasticity).
This means that:
P(t)
contains the consequences of previous experience.
A new input changes its local geometry, its active regions and its transition probabilities.
8.18 The inner world as an active model
The internal representation is not a passive copy of the surrounding world.
It is an active generative state.
For example, the present model may contain the expectations:
object behind obstacle,
person continues walking,
sound source remains present,
own body remains in position.
Such information need not be directly present in the sensory input at every moment.
8.19 The difference between a sensor and a percept
A sensor provides:
evidence.
A percept constitutes:
interpretation conditioned by current internal state.
Therefore:
same sensory evidence
may lead to a:
different percept
if:
previous internal state differs.
This is one of the main predictions of this chapter.
8.20 Continuity of object identity
One of the problems of perception is to preserve the identity of an object when its sensory properties change.
An object may:
move,
rotate,
change illumination,
become partially occluded.
The input changes markedly.
Nevertheless the internal model may maintain the:
same object identity.
DPSH assumes that this continuity may be a function of the trajectory in the Perceptual Manifold, and not merely of the similarity of individual frames.
8.21 The trajectory of an object
Instead of:
object = static pattern
the representation may be:
M_object(t).
The motion of the object:
position_1
->
position_2
->
position_3
corresponds to a continuous trajectory inside the dynamic manifold.
Object identity may then be associated with the continuity of this trajectory.
8.22 Prediction of motion
If the system knows:
position(t)
velocity(t),
it may predict:
position(t + dt).
If sensory data are briefly missing, the internal trajectory may continue:
predicted state evolution.
After the return of the sensory input, the prediction is compared with reality.
8.23 Correction versus reconstruction
Two possible strategies thereby arise.
Reconstructive strategy
current input
->
rebuild entire representation.
Corrective strategy
previous internal state
->
prediction
+
sensory error
->
corrected internal state.
DPSH expects that the second strategy corresponds better to a continuous perceptual system.
8.24 Hysteresis and attention
Attention may change the stability of certain regions of the manifold.
For example:
attention(A)
->
increase stability(M_A).
This may increase the hysteresis for the relevant percept.
Another state may conversely be left more easily.
Attention therefore need not create the percept.
It may change its dynamic stability.
8.25 Hysteresis and valuation
Evaluative systems may deform the dynamic landscape in a similar way.
If:
M_threat
is associated with a high value of relevance,
it may be:
easier to enter
harder to leave.
This provides a possible mechanism by which emotions and significance change perception.
8.26 Hysteresis and intuition
Long-term experience may create state biases.
If the system has learned in the past that a certain structure often precedes danger:
context X
->
M_warning.
With a new partial input, owing to hysteresis and the learned geometry:
partial X
->
rapid persistence of M_warning.
The action system may respond before the explicit reason is globally available.
This is compatible with the working interpretation of intuition.
8.27 Hysteresis as a source of expectation
The state of the system contains an implicit expectation of further development.
If the system remains:
S(t) in M_A,
then the probabilities of future states are:
P(M_j | M_A).
The past is thus projected into the future through the transition structure.
8.28 The temporal depth of the percept
A percept need not represent only the present.
It may contain:
trace of past,
current state,
prediction of future.
Provisionally:
percept(t)
=
F(
recent history,
current evidence,
expected continuation
).
A temporally deeper internal representation thereby arises.
8.29 The "present moment" as a temporal window
DPSH allows the hypothesis that the internal perceptual "now" is not a mathematical point in time.
It may be a dynamic interval in which:
recent past
still influences:
current state
and the present state already contains:
short-term prediction.
That is:
perceptual present
=
temporally extended dynamic state.
This claim is so far theoretical and requires separate experimental verification.
8.30 Continuity and the Global Workspace
If the Global Workspace gains access to:
M_A,
it need not receive merely a snapshot.
It may gain access to a dynamic state that already contains:
history,
current context,
expected transitions.
The workspace therefore need not reconstruct the temporal continuity itself.
It may take it over from the perceptual dynamics.
8.31 Global access is not a global reset
The broadcast of a state into the workspace should not cause:
reset perceptual dynamics.
On the contrary, the workspace may modulate, in return:
attention,
prediction,
action,
memory,
and thereby further deform this same ongoing internal state.
A closed loop arises:
perceptual manifold
->
workspace
->
modulation
->
perceptual manifold.
8.32 Experiment HY1 – hysteresis stimulus sweep
We create a continuous stimulus:
x ∈ [0,1]
which gradually passes between two interpretations:
A
B.
The first run:
x: 0 -> 1.
The second:
x: 1 -> 0.
We measure:
transition threshold.
If:
θ_AB != θ_BA,
the network exhibits hysteresis.
8.33 Experiment HY2 – short-term occlusion
The network creates:
M_object.
Then:
sensory object input = 0
for the duration:
Δt.
We track:
state persistence
and
prediction of reappearance.
We vary the length of the occlusion and look for the time over which the internal representation remains functional.
8.34 Experiment HY3 – conflict duration
After the formation of:
M_A
we begin to provide evidence for:
B.
We measure the time:
T_switch
needed for:
M_A -> M_B.
We track the relation:
prediction error strength
vs
transition time.
8.35 Experiment HY4 – state reset control
We compare two variants.
Continuous network
previous state preserved.
Reset network
Before every new input:
S -> S_baseline.
If continuity provides a functional advantage, we expect differences in:
ambiguous input interpretation,
occlusion handling,
temporal prediction,
object continuity,
generalization.
8.36 Experiment HY5 – identical current input, different history
We create:
history A -> X
and:
history B -> X.
X is identical.
We measure:
internal state,
downstream action,
prediction.
The strong prediction:
S(X | history A)
!=
S(X | history B).
8.37 Experiment HY6 – persistence curve
After the formation of the percept we remove the supporting stimulus.
We measure:
Q(t)
where Q expresses the quality or decodability of the perceptual state.
We may obtain, for example:
rapid collapse,
exponential decay,
plateau + transition.
The shape of the persistence curve will provide information about the mechanism of state maintenance.
8.38 Experiment HY7 – contradictory evidence
After the formation of:
M_A
we present progressively stronger evidence for:
B.
We measure:
P(M_B | evidence strength).
If the state exhibits dynamic inertia, the transition should be nonlinear.
8.39 Experiment HY8 – hysteresis after learning
We carry out the same hysteresis experiment:
before learning
and:
after learning.
If experience changes the dynamic geometry, there may be a change in:
θ_AB,
θ_BA,
hysteresis width.
This provides direct evidence that learning changes the persistence properties of the Perceptual Manifold.
8.40 Experiment HY9 – oscillator dependence
We measure hysteresis with:
phase intact
and:
phase scrambled.
If temporal organization contributes to the stability of the percept, there may be a change in:
hysteresis width
or
state lifetime.
We thereby connect the chapter on oscillations with continuity.
8.41 Experiment HY10 – stochasticity dependence
In the same way we vary:
σ.
We track whether stochasticity:
facilitates escape from old state
or:
destabilizes state excessively.
An optimal regime may arise between:
rigidity
and:
instability.
8.42 Experiment HY11 – dynamic versus explicit memory
A control network will solve continuity by means of:
explicit memory variable.
The dynamic network will make use of:
metastable state.
We compare:
occlusion,
ambiguous input,
perturbation recovery,
generalization,
spontaneous transition.
The aim is not to show that dynamic memory is always better.
The aim is to determine whether it exhibits properties that simple explicit memory does not explain.
8.43 A metric of continuity
We may define:
C_persistence(Δt)
as the similarity of the internal perceptual state before and after the loss of input:
C_persistence(Δt)
=
similarity(
M_before,
M_after(Δt)
).
We thereby obtain a quantitative measure of continuity.
8.44 A metric of history dependence
For an identical input X after two histories:
S_A(X)
S_B(X)
we may define:
H_dep =
D(S_A(X), S_B(X)).
If:
H_dep ~ 0,
history has no significant influence.
If:
H_dep >> 0,
the present representation is history-dependent.
8.45 A metric of hysteresis
A simple metric:
H =
|θ_AB - θ_BA|.
For a more complex manifold one may measure the difference of whole transition functions:
H_dynamic =
D(
P_forward,
P_reverse
).
8.46 Falsification criteria
The strong hypothesis of hysteresis and continuity will be weakened if:
- the same present input creates the same internal state regardless of history,
- the percept is extinguished immediately upon a short-term loss of sensory evidence,
- resetting the network does not change performance in temporally dependent tasks,
- forward and reverse stimulus sweeps exhibit no measurable hysteresis,
- the previous percept does not influence the interpretation of an ambivalent input,
- learned experience does not change persistence or transition geometry,
- all the observed continuity can be explained by a simple explicit memory variable without the need for metastable dynamics.
In such a case the conception of a continuous Perceptual Manifold would have to be markedly weakened.
8.47 Chapter research hypothesis
We formulate the partial hypothesis H7:
H7 – Perceptual Continuity and Hysteresis Hypothesis
The internal perceptual state is not created independently from each current sensory input, but arises continuously by transformation of the previous dynamic state. As a consequence, the perceptual system exhibits history dependence, state inertia and hysteresis: the same current input may be interpreted differently depending on the trajectory that preceded it.
The stronger prediction:
If the Perceptual Manifold genuinely constitutes an ongoing internal model of the world, then perceptual information must persist briefly even upon a loss of direct sensory evidence, the present state must influence the interpretation of a subsequent ambivalent input, and the transition thresholds between percepts must depend on the direction of the system's previous development.
This hypothesis does not claim:
persistence = consciousness.
It claims:
previous internal state
+
current sensory evidence
->
current percept.
One of the main properties of DPSH thereby arises:
The system does not have merely a series of representations of the world. It has a continuously existing inner world that new experience modifies.
9. Predictive constraint on the dynamics
9.1 The problem of free internal dynamics
The preceding chapters assume a system that is:
globally clockless,
recurrent,
stochastic,
oscillatorily organized,
history-dependent,
metastable.
Such a system may create rich internal dynamics.
This alone, however, is not sufficient.
If the internal state could evolve entirely independently of the external world, the system might create:
self-consistent
but:
externally incorrect
representations.
The Dynamic Perceptual State Hypothesis therefore needs a mechanism that maintains the internal dynamics in sufficient agreement with sensory reality.
This mechanism is the predictive constraint.
9.2 The internal state as a generative model
The perceptual state is not understood merely as a reaction to the current input.
It may at the same time generate expectations of future sensory development.
Formally:
P(t + dt) = G(S(t))
where:
S(t)is the present internal state,Gis the predictive mechanism,P(t + dt)is the expected future sensory state.
The system therefore creates not only:
representation of now,
but also:
expectation of next.
9.3 Sensory evidence
At the next moment the actual sensory input arrives:
I(t + dt).
This can be compared with the prediction:
P(t + dt).
A deviation arises:
ε(t + dt)
=
I(t + dt) - P(t + dt).
This deviation does not necessarily constitute a single central global scalar.
It may be distributed among many local subsystems:
ε_visual,
ε_audio,
ε_motion,
ε_position,
...
Each of them may locally change the dynamics of the network.
9.4 Prediction is not a central decision-maker
DPSH does not assume a mechanism:
global_error
->
choose_correct_percept.
Such a model would merely reintroduce a central selector.
Instead, the prediction error is to act locally.
For example:
local mismatch
->
altered excitability
->
altered spike probability
->
destabilized local pattern
->
changed recurrent dynamics.
A global change of the percept then arises as a macroscopic consequence of many local corrections.
9.5 Prediction as a constraint
The basic role of predictive processing in DPSH is not:
construct percept.
It is:
constrain perceptual dynamics.
That is:
possible internal states
+
sensory evidence
->
allowed / disfavored trajectories.
Prediction changes the probability that the system remains in a certain metastable region.
For a state M_A we may write:
stability(M_A)
=
F(
internal dynamics,
sensory consistency,
prediction error
).
9.6 Stabilization of a compatible state
If:
prediction(M_A)
≈
sensory evidence,
then:
ε_A ≈ 0.
In such a case:
persistence(M_A) high.
may hold.
The system therefore need not create a new percept at every new input.
The present state may continue if it remains compatible with reality.
9.7 Destabilization of an incompatible state
If:
prediction(M_A)
!=
sensory evidence,
then:
ε_A grows.
This may reduce:
stability(M_A).
If the deviation persists or is sufficiently large:
M_A
->
transition region
->
M_B.
The prediction error may therefore be one of the mechanisms that allow an old percept to be left behind.
9.8 Persistence versus correction
DPSH needs a balance between two tendencies.
Persistence
previous state
->
continuity.
Correction
sensory mismatch
->
state revision.
Too strong a persistence:
internal model ignores reality.
Too strong a correction:
internal model collapses with every small fluctuation.
A functional system must exist between these extremes.
9.9 Prediction error as pressure, not instruction
It is useful to understand the prediction error not as an instruction:
"switch to state B".
Rather as a pressure:
"state A is becoming less dynamically viable".
The system itself, through its internal dynamics, searches for a different configuration.
That is:
prediction error
->
destabilization
->
increased state-space exploration
->
new state selection.
This fits well in combination with stochasticity.
9.10 Stochasticity and predictive constraint
Spontaneous stochasticity may generate alternative trajectories:
M_A
->
candidate state M_B
->
candidate state M_C.
The predictive mechanism may change their relative stability.
For example:
ε_A = high
ε_B = low
ε_C = medium.
Then the following may arise:
P(M_A persists) low
P(M_B persists) high
P(M_C persists) medium.
Sensory reality therefore does not tell the system directly which state it is to create.
It changes the probability of survival of the available states.
9.11 Selection among internal hypotheses
Perceptual dynamics can be understood as an ongoing competition among alternative internal hypotheses:
H_A
H_B
H_C.
Each corresponds to a certain region of the manifold:
M_A
M_B
M_C.
Sensory evidence continuously changes their dynamic support.
This leads to:
hypothesis competition
->
metastable selection.
9.12 Predictive processing and symmetry breaking
If two interpretations have similar support:
M_A ~ M_B,
then a small difference in prediction error may create a bias:
ε_A < ε_B.
This difference may be amplified by recurrence:
small predictive advantage
->
larger state stability
->
recurrent reinforcement
->
symmetry breaking.
Prediction may thus influence the selection of the percept without an explicit selector.
9.13 Predictive processing and hysteresis
The previous percept may be stable even with a mild increase of the prediction error.
This creates hysteresis:
M_A persists
as long as:
ε_A < threshold_A.
Once a certain boundary is crossed:
M_A -> M_B.
In the opposite direction the return threshold may be different.
This provides a natural connection between:
prediction,
persistence,
hysteresis.
9.14 Prediction and temporal depth
The internal model need not predict only the next moment.
It may create expectations at several temporal scales:
short-term prediction,
medium-term expectation,
long-term context.
For example:
immediate sensory continuation,
object trajectory,
expected event sequence.
DPSH therefore admits a hierarchy of temporal predictions.
9.15 Local predictors
Prediction need not exist in a single central module.
We may have:
predictor_visual,
predictor_motion,
predictor_audio,
predictor_body,
predictor_context.
Each works with a locally available part of the dynamic state.
Their interaction may jointly constrain the global Perceptual Manifold.
9.16 Prediction and oscillation
A predictive signal may change:
phase,
amplitude,
excitability,
synaptic gain.
For example, an expected input may shift the local phase so that a certain population is more excitable at the moment of the expected signal.
A mechanism thereby arises:
prediction
->
temporal preparation
->
selective sensory gain.
9.17 Prediction as phase preparation
If the system expects an event at time:
t_expected,
a local oscillation may set:
phase(t_expected) = receptive phase.
An incoming expected spike then has a greater effect.
An unexpected input may arrive in a less advantageous phase and create a greater local disruption.
This connects predictive processing with temporal organization.
9.18 Prediction and synaptic plasticity
The prediction error may modulate plasticity.
For example:
high error
->
increased plasticity
or conversely:
high confidence
->
reduced plasticity.
In general:
Δw_ij
=
G(
spike timing,
local phase,
prediction error,
current state
).
An experience that surprises the system may thereby change the network differently than an expected event.
9.19 Learning a model of the environment
If the sequence:
A -> B -> C,
appears repeatedly, the network may change its connectivity so that:
A
increases the probability of the prediction:
B,
and:
B
increases the expectation of:
C.
The internal dynamics thereby becomes a generative model of the temporal structure of the environment.
9.20 The predictive manifold
The Perceptual Manifold is therefore not merely a map of present percepts.
It also contains a transition structure:
P(M_j | M_i).
This can be interpreted as prediction.
For example:
M_person_walking
->
high probability(M_person_next_position).
The internal manifold therefore contains an implicit model of:
what tends to happen next.
9.21 Prediction as the geometry of transitions
A prediction need not be represented by an explicit number.
It may be encoded geometrically.
If the transition:
M_A -> M_B
is dynamically very easy,
this may mean:
B is strongly expected after A.
If the transition is very improbable:
B is unexpected after A.
Prediction may thereby become a property of the state space itself.
9.22 Prediction without an explicit predictor
This leads to an important possibility.
It is not necessary that the network contain a component:
predict(B).
The dynamics itself may cause:
S in M_A
->
trajectory naturally moves toward M_B.
The prediction is then implicitly contained in the dynamic geometry.
9.23 Explicit and implicit prediction
DPSH therefore distinguishes:
Explicit prediction
A separate module creates:
predicted signal.
Implicit prediction
The topology and dynamics cause:
M_A -> likely M_B.
Both mechanisms may exist in Cognia.
It is important to determine experimentally which of them is necessary for a given task.
9.24 Prediction and non-commutativity
If:
A -> B
is the expected sequence,
but:
B -> A
is not,
then:
prediction_error(A -> B)
!=
prediction_error(B -> A).
The predictive mechanism is therefore naturally sensitive to order.
This connects predictive processing with non-commutative dynamics.
9.25 Prediction and causality
The system may learn not only correlations, but also directional relations:
A precedes B.
This by itself, however, does not yet demonstrate genuine causality.
DPSH therefore uses the notion:
temporal predictive relation
more cautiously than:
causal model.
A causal representation would require separate experiments with interventions.
9.26 Prediction and surprise
An unexpected event may have a special functional significance.
If:
ε >> normal range,
the system may:
increase attention,
increase plasticity,
destabilize current percept,
trigger workspace access.
Surprise may thereby become a mechanism of transition between a local perceptual state and broader global processing.
9.27 Prediction and the Global Workspace
One possible architecture:
predictable input
->
handled locally.
Conversely:
large unresolved prediction error
->
local instability
->
workspace access
->
global processing.
This would mean that the Global Workspace is not activated at every percept.
It may be especially important when the local model is not sufficient.
9.28 Prediction and intuition
An intuitive decision may arise if the learned manifold rapidly passes into the state:
M_warning
on the basis of a partial input.
This transition may be the result of a long-term learned predictive geometry.
The system need not know explicitly:
"which exact features caused the warning".
Nevertheless the state:
M_warning
may correctly predict:
unfavorable outcome.
9.29 Prediction and phenomenal stability
If a subjective percept is stable despite sensory noise, one of the explanations may be precisely the dominance of internal prediction over small instantaneous deviations.
This must not, however, be formulated as proof of the relation:
predictive processing -> qualia.
DPSH uses the predictive mechanism only as a candidate principle of the stability of the functional percept.
9.30 Hallucination as an extreme of internal dominance
A hypothetical pathological regime can be described as:
internal prediction influence
>>
sensory correction.
Then:
self-generated state
->
self-confirming dynamics.
Such a state may persist even with insufficient sensory support.
DPSH does not put forward a model of clinical hallucinations.
This extreme merely shows why the balance between the internal state and external evidence must be experimentally controlled.
9.31 Sensory chaos as the opposite extreme
At the opposite side:
sensory correction
>>
internal persistence.
Then every small change of input fundamentally rebuilds the state.
The system may lose:
continuity,
object permanence,
robust interpretation.
Functional perception may therefore require an intermediate regime.
9.32 Precision weighting
Not all sensory errors need have the same significance.
The system may estimate the trustworthiness of an input:
precision.
Then:
effective_error
=
precision * prediction_error.
A very noisy sensor, for example, may have:
low precision.
Its deviation will not destabilize the percept as much as a precise input.
9.33 Precision as modulation of the dynamics
Precision need not be an explicit probability.
It may be realized, for example, by:
gain modulation,
excitability,
synaptic strength,
oscillatory coupling.
Certain sensory channels thereby become more significant in a given context.
9.34 Attention as precision control
One of the possible interpretations of attention is:
attention
->
increase precision of selected signals.
In DPSH this would mean:
selected evidence
->
stronger influence on manifold dynamics.
Attention may thus change which prediction errors have the greatest influence on the stability of the percept.
9.35 Hierarchical predictions
A perceptual system may have several levels.
For example:
edges
->
shapes
->
objects
->
scenes.
Every level may predict the lower one.
At the same time it may receive a prediction error in the upward direction.
DPSH is not dependent on one particular predictive hierarchy, but it must be compatible with the possibility of multilevel dynamics.
9.36 Prediction and the Perceptual Manifold at several scales
We may have local manifolds:
P_visual,
P_audio,
P_body
and a broader one:
P_global.
Predictions may act:
within level
as well as:
across levels.
A global percept may arise as the coordination of several dynamic spaces.
9.37 Experiment P1 – expected versus unexpected continuation
We train the network on the sequence:
A -> B.
We then test:
A -> B
and:
A -> C.
We measure:
internal perturbation,
prediction-error activity,
transition time,
state stability.
We expect a greater dynamic disruption for:
A -> C.
9.38 Experiment P2 – prediction under occlusion
The network tracks a moving object:
x1 -> x2 -> x3.
The object then disappears for a while.
The network must internally predict:
x4,
x5.
After its return we measure the difference:
predicted position
vs
actual position.
This tests whether the Perceptual Manifold contains a dynamic prediction, and not merely a static memory of the last input.
9.39 Experiment P3 – contradictory evidence
We first create:
M_A.
We then progressively increase the sensory evidence for:
B.
We measure:
prediction error,
dwell time(M_A),
transition threshold,
transition trajectory.
This directly connects prediction error with hysteresis.
9.40 Experiment P4 – prediction error ablation
We create two identical networks.
Network A
Prediction-error feedback active.
Network B
Prediction-error feedback removed or markedly limited.
We test:
percept stability,
adaptation,
hallucination-like persistence,
response to changed environment.
If both networks function in the same way, the role of the predictive constraint will be weakened.
9.41 Experiment P5 – excessive prediction gain
We will increase the strength of the top-down prediction:
g_pred.
We measure:
sensory correction,
persistence,
mismatch tolerance.
We expect that too high a value may lead to:
excessive persistence.
9.42 Experiment P6 – excessive sensory gain
Conversely we increase:
g_sensory.
We measure:
state stability,
sensitivity to noise,
percept continuity.
Too high a value may cause:
unstable frame-like perception.
9.43 Experiment P7 – optimal prediction/sensory balance
We will explore a two-dimensional space:
g_prediction
x
g_sensory.
We look for the region where the system simultaneously attains:
persistence,
adaptability,
prediction accuracy,
robustness.
9.44 Experiment P8 – precision modulation
We add noise to the sensor.
The network receives the information, or must estimate:
sensor reliability.
We test whether it is able to reduce the influence of an unreliable error signal.
We compare a:
precision-aware
and a:
precision-unaware
system.
9.45 Experiment P9 – implicit versus explicit prediction
We compare:
Model A
An explicit predictor unit.
Model B
Prediction only as a learned transition geometry.
Both systems solve the same temporal task.
We measure:
accuracy,
robustness,
generalization,
state-space structure.
In this way we find out whether we need an explicit predictor at all.
9.46 Experiment P10 – surprising event and workspace candidate
The network runs in a stable:
M_A.
We insert an unexpected event:
X_surprise.
We measure:
prediction error,
state destabilization,
activation of global-access mechanism.
Later it will be possible to test the hypothesis:
unresolved surprise
->
increased probability of workspace ignition.
9.47 Experiment P11 – prediction changes the phase structure
We test whether the expectation of a certain input changes, before its arrival:
oscillatory phase,
phase coherence,
local excitability.
If so, it is possible to connect directly:
prediction
->
temporal preparation.
9.48 A metric of prediction error
For a simple experiment:
E_pred =
D(P(t), I(t)).
For a complex manifold it may be more appropriate to measure the difference between:
expected population state
and:
observed sensory-driven state.
9.49 A metric of predictive stability
We define:
R_pred(M_A)
as the relation between the prediction error and the probability of leaving the state:
R_pred =
P(exit M_A | error ε).
If the predictive constraint works, this function should be systematic.
9.50 A metric of adaptability
The system must respond to a genuine change of the environment.
We may measure:
T_adapt
from the moment of a permanent change of input to the stabilization of the new state.
Too high a value:
excessive rigidity.
Too low:
insufficient persistence.
9.51 A metric of model consistency
If the internal state genuinely represents the structure of the environment, its predictions should be statistically better than baseline.
That is:
prediction_accuracy
>
naive predictor.
Without this property it would not be legitimate to speak of a generative internal model.
9.52 Deep State Learning and prediction
The stronger hypothesis of Deep State Learning assumes that the prediction error may shape not only the output, but the geometry of the internal manifold itself.
That is:
repeated prediction error
->
plasticity
->
new state-space geometry
->
improved future prediction.
The network thereby learns:
structure of possible worlds
not merely:
correct output labels.
9.53 Learning transition probabilities
If the environment exhibits:
A -> B 80 %
A -> C 20 %,
then after learning we expect:
P(M_B | M_A)
>
P(M_C | M_A).
The dynamics of the manifold should reflect the statistics of experience.
9.54 Spontaneous activity and predictive replay
After the removal of the external input, spontaneous activity may visit the learned trajectories:
M_A -> M_B -> M_C.
If plasticity is taking place at the same time, the system may further change its model.
Here a key question arises:
consolidation
versus:
self-reinforcing error.
This boundary will have to be measured experimentally.
9.55 Prediction as protection against self-reinforcement
If internal replay creates a state that is not subsequently confirmed by external experience, the prediction error may prevent its uncontrolled amplification.
This may be one of the mechanisms that separates:
useful internal learning
from:
self-generated drift.
9.56 Falsification criteria
The predictive part of DPSH will be weakened if:
- the internal state does not generate measurably better predictions of future input than baseline,
- the prediction error does not change the stability of metastable states,
- removing the prediction-error feedback does not change adaptation to the environment,
- during occlusion the system merely holds the last value and does not generate a dynamic prediction,
- the learned transition probabilities do not correspond to the statistics of the environment,
- top-down prediction provides no advantage with a noisy or ambivalent input,
- the predictive mechanism can be removed without a change of the relevant perceptual properties,
- all the effects can be explained by a simple feed-forward correction without a relation to the internal state space.
In such a case, predictive processing will not be a central constraint of DPSH.
9.57 Chapter research hypothesis
We formulate the partial hypothesis H8:
H8 – Predictively Constrained Dynamics Hypothesis
Metastable perceptual dynamics is continuously constrained by the difference between predicted and actual sensory evidence. States compatible with the environment have a higher probability of persistence, while persistently incompatible states are destabilized. Prediction thus does not determine the percept centrally, but changes the dynamic geometry and the probabilities of transitions between internal states.
The stronger prediction reads:
If the internal state genuinely functions as a generative model, it must predict the future sensory structure better than a simple statistical baseline, and systematic manipulation of the prediction-error feedback must change the stability, adaptability and transition geometry of perceptual states.
This hypothesis does not claim:
prediction = perception
nor:
prediction = consciousness.
It claims:
autonomous internal dynamics
+
sensory prediction
+
distributed prediction error
->
reality-constrained perceptual manifold.
A key property of DPSH thereby arises:
The inner world need not be reconstructed anew from every sensory input. It may exist continuously as its own dynamic model, provided it is continuously corrected by actual experience.
10. Learning the dynamics and Deep State Learning
10.1 Learning is not merely a mapping of input onto output
In ordinary machine learning the aim is often to find a transformation:
input
->
output
such that the resulting output minimizes a certain error.
A typical model therefore optimizes the relation:
Y = F(X).
The Dynamic Perceptual State Hypothesis, however, assumes a system whose functional significance does not lie solely in the resulting output.
What matters is the dynamics itself:
S(t0)
->
S(t1)
->
S(t2)
->
...
Learning therefore need not mean only a change of:
input -> output mapping.
It may also mean a change of:
state-space geometry,
transition probabilities,
metastable state stability,
temporal organization,
predictive structure.
The working term Deep State Learning designates precisely this second level of learning.
10.2 What Deep State Learning means
Deep State Learning here does not designate a deep neural network in the sense of a large number of layers.
The word deep designates the fact that learning takes place inside the
dynamic state space of the system.
Instead of:
learn correct output
the system learns:
which internal states should exist,
which states should be stable,
which transitions should be probable,
which temporal relations should be reinforced,
which predictions should follow from which states.
Schematically:
experience
->
internal trajectory
->
local plasticity
->
modified future trajectories.
10.3 Learning as a change of the dynamic geometry
Let us consider a network with the dynamics:
dS/dt = F(S, I, W).
The synaptic parameters:
W
determine part of the geometry of the state space.
If learning changes:
W -> W',
the dynamics changes as well:
F -> F'.
This changes the:
basins,
trajectories,
transition probabilities,
metastable states.
Learning can therefore be interpreted as a deformation:
Ω_before
->
Ω_after.
10.4 Experience changes future possibilities
If the system repeatedly experiences the sequence:
A -> B -> C,
then after learning it need not merely classify correctly:
A,
B,
C.
The dynamics itself may change so that:
M_A -> M_B
is more probable than:
M_A -> M_X.
Similarly:
M_B -> M_C
may become the preferred trajectory.
Experience thereby changes:
future possibilities.
10.5 Local plasticity
DPSH prefers mechanisms that can be realized locally.
A synapse:
w_ij
need not know the global state of the network.
Its change may depend, for example, on:
presynaptic spike,
postsynaptic spike,
relative timing,
local phase,
modulatory signal,
prediction error,
current synaptic state.
In general:
Δw_ij =
G(
t_pre,
t_post,
φ_local,
ε_local,
w_ij,
state_i,
state_j
).
10.6 STDP as a basic time-sensitive mechanism
Spike-timing-dependent plasticity allows:
Δw_ij =
F(t_post - t_pre).
The order:
pre -> post
may therefore lead to a different change than:
post -> pre.
STDP therefore naturally connects:
timing
and:
learning.
Within DPSH, however, STDP is not an end in itself.
It is one of the mechanisms by which temporal organization may be written into the future dynamics of the network.
10.7 Phase as a modulator of learning
If spike timing depends on a local oscillation, then:
oscillatory phase
->
spike timing
->
plasticity.
A stronger model may contain directly:
Δw_ij =
G(
Δt,
φ_pre,
φ_post
).
This means that the same pair of spikes need not have the same learning effect in different dynamic contexts.
10.8 Delay as a learnable temporal structure
Synaptic delays:
d_ij
change the moment at which information reaches its target.
If plasticity selectively strengthens paths whose signals arrive at a functionally suitable moment, the network may learn not only:
who connects to whom,
but also:
which temporal routes matter.
This leads to the possibility that learning is written into:
connectivity
+
effective timing.
10.9 Learning the temporal topology
After repeated experience there may arise:
path A
->
preferred timing
->
strengthened route.
Other paths may weaken.
The result is a learned temporal topology:
T_W.
This determines:
when information can efficiently propagate.
10.10 Recurrence as an object of learning
In a recurrent network, plasticity does not change only the forward passage of information.
It also changes the:
feedback loops,
recurrent gain,
attractor geometry,
metastability.
A small local change may therefore gradually change the global dynamic regime.
10.11 Learning a metastable state
Before learning, a certain configuration may exist only briefly:
M_A
->
rapid decay.
After repeated experience, plasticity may create:
stronger recurrent support.
Then:
lifetime(M_A) increases.
Learning may therefore stabilize a perceptual region without necessarily turning it into a permanent attractor.
10.12 Learning the boundaries between states
Just as important as the stability of the state itself are the boundaries between:
M_A
M_B.
Plasticity may change the:
transition threshold,
basin size,
perturbation sensitivity.
Experience may thus change the way in which the system distinguishes between similar percepts.
10.13 Learning as the creation of a preference
If the network repeatedly sees:
A -> B,
there may arise:
P(M_B | M_A) high.
If:
A -> C
occurs rarely:
P(M_C | M_A) low.
The system thereby learns the statistical structure of the environment.
10.14 Learning and predictive processing
The prediction error provides one of the candidate modulatory signals.
If:
prediction is correct,
only a small synaptic change may be needed.
If:
prediction error is high,
there may be an increase in:
local plasticity.
In general:
plasticity_gain =
F(prediction_error).
10.15 Prediction error as a learning constraint
It is important not to assume a:
global optimizer.
The prediction error may only locally change the probability of a synaptic change.
For example:
local unexpected event
->
increased local plasticity.
A global improvement of the model then arises out of many such local changes.
10.16 Learning the internal model
If the system repeatedly predicts:
B
after the state:
A
and this prediction is confirmed, the dynamics:
M_A -> M_B
may be strengthened.
If the prediction repeatedly fails:
M_A -> M_B
may be weakened.
The internal generative model is thus written into the dynamics of the manifold.
10.17 Spontaneous activity as an internal training signal
A special part of DPSH is the hypothesis that learning need not take place only during external input.
If:
I_external = 0,
the network may still generate:
spontaneous trajectories.
For example:
M_A -> M_B -> M_C.
If plasticity remains active during this, these internally generated trajectories may further change:
W.
10.18 Ongoing learning
A mechanism thereby arises:
external experience
->
learned structure
->
spontaneous replay
->
further plasticity
->
modified internal structure.
We designate this process as:
ongoing learning.
It is one of the strongest parts of the Deep State Learning hypothesis.
10.19 Reactivation without an external teacher
During spontaneous replay the system has no new external teaching signal.
Nevertheless it may activate:
previously learned assemblies,
transitions,
temporal sequences.
This opens up the possibility of:
consolidation
without new experience.
10.20 Consolidation
A useful regime may look like:
experience
->
weak learned structure
->
replay
->
repeated local timing
->
strengthened useful transitions.
A structure that arose in the short term may thereby become more stable.
10.21 Generalization
Spontaneous dynamics need not reproduce earlier experience exactly.
Stochasticity may generate:
variations around learned states.
If plasticity preserves the common relations and removes the random details, there may arise:
generalization.
For example:
A1,
A2,
A3
may gradually create a broader:
M_A.
10.22 The danger of self-reinforcement
The same mechanism may, however, have the opposite consequence.
If the network spontaneously creates an erroneous trajectory:
M_X,
and plasticity automatically strengthens it:
M_X
->
stronger probability of M_X,
there arises:
self-reinforcement.
The system may amplify its own errors.
10.23 Internal hallucination as a learning failure
An extreme case:
spontaneous state
->
plasticity
->
stronger spontaneous state
->
further plasticity.
Without a sufficient external constraint the network may create internal structures that are not supported by the environment.
Deep State Learning therefore needs a control mechanism.
10.24 Sensory evidence as a correction of learning
The external world provides a corrective signal.
If an internally strengthened state repeatedly generates an erroneous prediction:
prediction error high,
it should be:
destabilized
or
relearned.
This closes the loop:
internal learning
<->
external validation.
10.25 Learning gate
One of the possibilities is to modulate when plasticity is permitted.
For example:
plasticity =
F(
prediction confidence,
novelty,
reward,
attention,
internal state
).
Not every spike therefore has to change a synapse automatically.
10.26 Consolidation gate
Spontaneous replay may be useful only in some states.
We may therefore introduce:
consolidation mode.
In it:
spontaneous activity
+
selected plasticity.
This may later be compared experimentally with:
always-on plasticity.
10.27 Homeostatic plasticity
So that the network does not converge to extreme weights, it may need homeostatic mechanisms.
For example:
target firing range,
weight normalization,
inhibitory balancing.
Homeostasis is not the main principle of consciousness in DPSH.
It may, however, be a necessary stabilizing mechanism of the research architecture.
10.28 Plasticity of excitation and inhibition
Learning need not change only excitatory synapses.
The inhibitory structure may be equally important for:
competition,
state boundaries,
oscillations,
metastability.
It must therefore be possible to separate experimentally:
excitatory plasticity
and:
inhibitory plasticity.
10.29 Learning the oscillatory organization
If local rhythms are emergent, plasticity may also change their:
frequency,
coherence,
phase relations.
That is:
learning
->
temporal organization.
At the same time:
temporal organization
->
learning.
A bidirectional coupling arises.
10.30 Learning the phase geometry
For local oscillators:
O1 ... Om
we may track:
Φ(t).
After experience the probability of certain configurations may change:
P(Φ).
Learning therefore need not create only neuronal assemblies.
It may create preferred phase relations between them.
10.31 Learning dynamic routing
If communication efficiency depends on the relative phase:
C_ij = F(Δφ_ij),
the network may by learning change the:
phase relations
so that certain communication paths are more easily available.
The result is a learned dynamic routing.
10.32 Deep State Learning and non-commutativity
Because the order of events changes plasticity:
A -> B
!=
B -> A,
the system learns directional sequences.
The resulting dynamics therefore contains the history of experience.
Deep State Learning is not merely the statistics of occurrence.
It is also the statistics of order.
10.33 Deep State Learning and hysteresis
If experience changes the basins and their boundaries, it may also change the:
hysteresis width.
For example, after repeated use of the state:
M_A
it may be:
easier to enter,
harder to leave.
Learning may thus manifest itself as a change of the dynamic inertia of the percept.
10.34 Deep State Learning and intuition
After long learning, the path:
input X -> M_A
may be very short and dynamically preferred.
The system may then rapidly arrive at the relevant macrostate without an explicit reconstruction of all the learned rules.
Such a mechanism may provide a functional basis for intuitive decision-making.
10.35 Explicit knowledge versus implicit geometry
A system may have knowledge represented explicitly:
rule:
if X then Y.
Or implicitly:
geometry causes:
X -> M_Y.
The second form need not be directly available for a symbolic report.
Nevertheless it may influence decision-making.
10.36 Learning Global Workspace access
Later it may also be learnable:
which perceptual states gain workspace access.
For example, states with high:
novelty,
uncertainty,
relevance,
prediction error
may more often gain global availability.
This part, however, belongs to the next chapter.
10.37 Multiple temporal scales of learning
Synaptic changes may have different time constants.
For example:
fast plasticity
->
short-term adaptation.
slow plasticity
->
long-term state-space geometry.
DPSH allows the existence of several learning timescales.
10.38 The short-term plastic trace
Some changes may decay rapidly:
Δw_fast(t).
They may temporarily influence the:
current percept,
recent context,
transition probability.
This provides a further mechanism between pure activity and long-term memory.
10.39 Long-term learning
Repeatedly confirmed relations may pass into:
Δw_slow.
Experience thereby gradually becomes part of the long-term geometry of the Perceptual Manifold.
10.40 Consolidation across temporal scales
A possible architecture:
fast state
->
repeated replay
->
slow synaptic consolidation.
Such a mechanism would allow:
experience now
to be gradually converted into:
persistent future bias.
10.41 Deep State Learning is not a backpropagation requirement
DPSH does not require the network to learn by means of global backpropagation.
On the contrary, the main hypothesis assumes that the relevant dynamic structures may arise through:
local plasticity,
local timing,
local error modulation,
recurrent interaction.
This is an experimental assumption, not a dogma.
If local mechanisms turn out not to be sufficient, the hypothesis will have to be revised.
10.42 Controller versus local learning
Cognia may contain a controller.
It is important, however, to separate:
controller decides what to learn
from:
controller computes global desired weights.
The first possibility is compatible with DPSH.
The controller may modulate:
learning on/off,
attention,
relevance,
plasticity gain.
The synaptic change itself may, however, remain local.
10.43 Internal relevance
Not all experiences need have the same learning strength.
The system may generate an internal value:
relevance.
This may be a function of:
surprise,
reward,
threat,
novelty,
prediction error.
Then:
learning_rate =
F(relevance).
10.44 Learning without conscious access
If local dynamics can change synapses before entry into the Global Workspace, there may take place:
implicit learning.
The system may therefore learn the structure of the environment without the whole learning process being globally available.
10.45 Conscious learning as a modulated regime
The Global Workspace may conversely enable:
top-down attention,
deliberate replay,
explicit strategy.
This may strengthen or direct Deep State Learning.
At least two levels therefore arise:
implicit state-space learning
and:
globally modulated learning.
10.46 What exactly is to be measured in Cognia
It is not sufficient to track:
final accuracy.
For Deep State Learning it is necessary to record:
weight distributions,
delay distributions,
phase relationships,
metastable state count,
state lifetime,
transition matrices,
basin geometry,
spontaneous/evoked similarity,
prediction accuracy,
generalization,
representational drift.
10.47 Experiment D1 – state space before/after learning
We measure:
Ω_before.
We expose the network to a structured environment.
We then measure:
Ω_after.
We compare:
cluster structure,
trajectories,
transition probabilities,
state stability.
If learning takes place at the level of the dynamics, the change must be observable.
10.48 Experiment D2 – sequence learning
We train the network on:
A -> B -> C.
We then test:
A.
We measure whether the dynamics spontaneously heads towards:
M_A -> M_B -> M_C.
The control sequence:
A -> C -> B
should have a lower probability if it was not learned.
10.49 Experiment D3 – timing-dependent learning
We use the same events, but a different order:
A -> B
versus:
B -> A.
After learning we compare:
connectivity,
transitions,
future prediction.
In this way we test whether non-commutative experience is written into the dynamics.
10.50 Experiment D4 – phase-dependent learning
We present the same sequence with:
phase structured
and:
phase scrambled.
We measure:
learning speed,
state stability,
transition geometry.
If the phase organizes learning, the difference must be measurable.
10.51 Experiment D5 – learning with plasticity OFF
A control network:
plasticity OFF.
We compare it with:
plasticity ON.
If the manifold does not change, the learning hypothesis fails.
10.52 Experiment D6 – STDP versus rate-based plasticity
We compare:
STDP
and:
matched rate-based Hebbian learning.
We measure not only task performance, but chiefly:
temporal structure,
transition directionality,
phase sensitivity,
metastability.
In this way we find out whether timing-sensitive plasticity adds anything specific.
10.53 Experiment D7 – spontaneous replay
After learning:
external input = 0.
We track whether the network spontaneously visits:
learned states.
We compare:
spontaneous before learning
spontaneous after learning.
We measure:
spontaneous/evoked similarity.
10.54 Experiment D8 – plasticity during replay
We create two networks:
A
spontaneous activity ON
plasticity ON.
B
spontaneous activity ON
plasticity OFF.
After a blank period we test:
memory,
prediction,
generalization,
state geometry.
This is a key test of ongoing learning.
10.55 Experiment D9 – silent control
A third variant:
C
spontaneous activity OFF
plasticity OFF.
The comparison:
A vs B vs C
will make it possible to separate:
effect of spontaneous dynamics
from:
effect of learning during spontaneous dynamics.
10.56 Experiment D10 – held-out generalization
After spontaneous learning we must not test only familiar patterns.
We use:
held-out stimuli.
If:
training performance improves
but:
held-out performance decreases,
this may be self-reinforcing memorization instead of a genuine improvement of the internal model.
10.57 Experiment D11 – self-reinforcement test
We deliberately insert into the network a small erroneous internal association:
A -> X_wrong.
We then allow spontaneous learning.
We track whether the:
wrong association decays,
remains,
amplifies.
This is a direct test of the stability of Deep State Learning.
10.58 Experiment D12 – external correction
After the internal erroneous association has been strengthened, we repeatedly present the correct sequence:
A -> B_correct.
We measure whether the prediction error is able to:
weaken wrong dynamics,
restore correct transition.
In this way we test the ability of the system to correct its own errors.
10.59 Experiment D13 – consolidation gate
We compare:
plasticity always ON
against:
plasticity active only under consolidation condition.
We measure:
stability,
drift,
generalization.
This may show whether ongoing learning needs to be controlled.
10.60 Experiment D14 – learning-rate sweep
We will vary:
η.
Too low:
no useful adaptation.
Too high:
instability / catastrophic drift.
We look for the region:
η*.
10.61 Experiment D15 – stochasticity × plasticity
We test the grid:
σ x η.
We look for the regime in which:
state exploration
+
stable learning
leads to the best generalization.
In this way we directly connect stochasticity with Deep State Learning.
10.62 Experiment D16 – oscillator × plasticity
Similarly:
phase structure ON/OFF
x
plasticity ON/OFF.
We measure:
learned temporal geometry.
If phase organization genuinely structures learning, the interaction effect must be measurable.
10.63 Experiment D17 – delay learning
If Cognia permits adaptive delays, we track:
d_ij before
d_ij after.
We test whether the network learns temporally compatible paths.
If delays are not learnable, it is at least possible to test the selection of weights according to fixed delays.
10.64 Experiment D18 – basin deformation
For a selected percept:
M_A
we map:
basin before learning.
After learning:
basin after learning.
We measure:
basin volume,
perturbation robustness,
entry probability.
In this way we directly test the metaphor of "learning as a deformation of the landscape".
10.65 Experiment D19 – hysteresis after learning
We measure:
H_before
and:
H_after.
If learning has stabilized a particular percept, the following may increase:
hysteresis width.
This connects Deep State Learning with continuity.
10.66 Experiment D20 – dynamic routing after learning
We measure the effective connectivity:
C_ij(t)
before learning and after it.
If phase relations have been learned, new preferred communication paths may arise without a change of the coarse anatomical topology.
10.67 A metric of manifold change
We define, for example:
D_manifold =
D(
P_before,
P_after
).
It must include more than one parameter.
For example:
state centroids,
transition matrix,
dwell times,
phase geometry.
10.68 A metric of spontaneous/evoked similarity
For a learned percept:
M_A
we compare:
evoked trajectory
and:
spontaneous trajectory.
We define:
R_replay =
similarity(
T_spontaneous,
T_evoked
).
A higher value after learning supports the reactivation of the learned dynamics.
10.69 A metric of generalization
The most important protection against self-reinforcement:
G =
performance(held-out data).
Deep State Learning is useful only if it does not merely increase internal confidence, but preserves or improves the ability to respond to new experience.
10.70 A metric of representational drift
During spontaneous learning we track:
D(
M_A(t0),
M_A(t1)
).
A slow adaptive change may be useful.
Uncontrolled drift:
D -> large
may mean the disintegration of the representation.
10.71 The stability-plasticity dilemma
The system must resolve a conflict:
plastic enough to learn
versus:
stable enough to remember.
DPSH expects that metastable dynamics, several learning timescales and local modulation of plasticity may provide mechanisms for this compromise.
This must, however, be verified experimentally.
10.72 What would be the strongest result
A very strong result would be:
- external experience creates distinguishable metastable states,
- after the removal of input the network spontaneously visits these states,
- plasticity during spontaneous activity changes their geometry,
- this change improves prediction or generalization on held-out data,
- the effect disappears when stochasticity, timing-sensitive plasticity or the relevant phase organization is switched off.
Such a result would be far stronger than a mere:
network remembers stimulus.
10.73 Falsification criteria
The Deep State Learning hypothesis will be weakened if:
- spontaneous activity merely reproduces the learned patterns without any further functional effect,
- plasticity during spontaneous activity does not improve any relevant property,
- ongoing learning worsens generalization,
- internal replay systematically amplifies its own errors,
- the state-space geometry does not change after learning,
- simple rate-based learning provides the same results,
- timing, phase and delays are not relevant for learning,
- all the necessary properties can be explained by explicit memory and supervised mapping without dynamic state-space learning.
10.74 Deep State Learning as a separately falsifiable part
The whole of DPSH need not fall if Deep State Learning fails.
It is possible that:
dynamic perceptual states exist
but:
spontaneous ongoing learning is not beneficial.
This hypothesis must therefore be tested separately.
10.75 Relation to phenomenal experience
Even successful Deep State Learning does not demonstrate the emergence of qualia.
It would, however, show an important property:
the internal state is not merely a passive working memory, but an active dynamic structure that itself influences its own future organization.
This would strengthen the functional model of continuous internal experience.
The phenomenal interpretation, however, remains a separate hypothesis.
10.76 Requirements for Cognia
For the testing of Deep State Learning, Cognia must allow at least:
- local timing-sensitive plasticity,
- modulation of plasticity by local signals,
- separate switching on/off of plasticity,
- spontaneous activity without external input,
- replay with plasticity switched on and off,
- several temporal scales of plasticity,
- recording of the history of synaptic changes,
- recording of the phase at learning events,
- synaptic delays,
- an experimental variant of rate-based plasticity,
- freeze and restore of the network state,
- exact comparison of the manifold before and after learning,
- held-out tests,
- the possibility of measuring representational drift,
- the possibility of controlling the learning gate by an external or internal modulator.
10.77 Chapter research hypothesis
We formulate the partial hypothesis H9:
H9 – Deep State Learning Hypothesis
Learning in a dynamic recurrent spiking network need not consist only in the mapping of sensory inputs onto outputs. Through local timing-sensitive plasticity, experience may change the geometry of the internal state space, the stability of metastable perceptual states and the probabilities of transitions between them. Spontaneous ongoing activity may reactivate these learned structures and, under certain conditions, enable their further consolidation or reorganization even without new external input.
The stronger falsifiable prediction reads:
If Deep State Learning genuinely exists as a functionally significant mechanism, then plasticity during internally generated spontaneous dynamics must cause a measurable change of the Perceptual Manifold that improves at least some properties of prediction, robustness or generalization on new data compared with an identical network with spontaneous activity but plasticity switched off.
The hypothesis therefore does not claim:
spontaneous activity automatically improves learning.
It claims:
spontaneous structured dynamics
+
appropriately constrained local plasticity
->
possible continued learning of internal state geometry.
The key question is not merely:
"Can the network learn something?"
But:
Can the network learn its own dynamic space so that its internal states correspond better to the structure of the world, and can it further consolidate this structure through its own internal activity?
11. Global Workspace and the global availability of the percept
11.1 The formation of a percept is not the same as global availability
The Dynamic Perceptual State Hypothesis distinguishes two different questions:
- how a coherent perceptual state arises,
- how this state becomes available to the wider system.
The preceding chapters dealt predominantly with the first question.
The working chain of DPSH is:
sensory dynamics
->
metastable perceptual state
->
workspace access
->
global availability.
In DPSH the Global Workspace is therefore not necessarily the mechanism that creates the percept.
It may be the mechanism that makes a selected, already existing dynamic state available to further functional subsystems.
11.2 A percept may exist without a global broadcast
Let us assume a local or distributed state:
M_A.
This state may be:
decodable,
persistent,
behaviorally relevant,
without having been made available to all the other modules.
It may, for example, influence:
local action selection
but it need not be available to:
language,
explicit report,
deliberate reasoning.
This creates an important distinction:
perceptual processing
versus:
globally accessible processing.
11.3 The Global Workspace as a functional interface
We provisionally define the Global Workspace as a mechanism that allows a selected state to influence several otherwise specialized subsystems.
For example:
perceptual state M_A
|
v
workspace
|
+-------+--------+--------+---------+
| | | | |
v v v v v
memory action value language planning.
The workspace therefore need not contain a complete copy of the whole Perceptual Manifold.
It may make available a certain projection of it or the currently relevant macrostate.
11.4 Specialized subsystems
DPSH assumes the existence of specialized functional modules.
For example:
visual processing,
auditory processing,
memory,
valuation,
motor control,
language,
planning,
attention.
Each of them may have its own local dynamics.
The Global Workspace provides a mechanism by which some results of these local dynamics may gain a wider influence.
11.5 Local availability
Not every state must be global.
There may be a hierarchy of availability:
local state
->
regional availability
->
workspace candidate
->
global broadcast.
We thereby avoid the notion that every neuronal state is automatically part of conscious processing.
11.6 A candidate for the workspace
A perceptual state may gain a higher probability of workspace access if it exhibits, for example:
high relevance,
novelty,
unresolved prediction error,
strong persistence,
behavioral importance,
competition victory,
attentional amplification.
Formally:
P(access | M_i)
=
F(
relevance,
novelty,
prediction_error,
stability,
attention,
context
).
11.7 Workspace selection need not be a central selector
Just as with perceptual selection, we do not wish to introduce a function:
workspace.select(M_A).
That would merely move the problem to another central mechanism.
Access may arise from the competition among candidate states:
M_A
M_B
M_C.
Each has a certain capacity to activate the distributed workspace network.
One representation may cross a dynamic threshold and trigger global propagation.
11.8 Ignition
The working mechanism may take the form:
local activation
->
recurrent amplification
->
threshold crossing
->
distributed ignition
->
global accessibility.
Ignition may therefore be a dynamic transition, not an explicit programmatic event.
11.9 Ignition as a phase transition
One possibility is to understand workspace ignition by analogy as a macroscopic phase transition.
Below a certain parameter:
λ < λ_c
activity remains local.
Above:
λ > λ_c
a qualitatively different regime arises:
distributed coherent activation.
This would correspond to:
local processing
->
global regime.
DPSH regards this possibility as testable, not as a settled conclusion.
11.10 Percept formation before ignition
One of the main distinguishing features of DPSH is the hypothesis:
perceptual state formation
precedes
workspace ignition.
That is:
M_A
may already contain a coherent perceptual structure before it becomes globally available.
The workspace then resolves:
accessibility,
not necessarily:
construction.
11.11 A stronger alternative
It is, however, necessary to admit the alternative as well:
Some percepts may arise only through broad recurrent interaction involving the workspace itself.
Experiments must therefore distinguish:
local percept formation
from:
workspace-dependent percept formation.
DPSH must not assume the result in advance.
11.12 The workspace as a broadcast of a dynamic state
If a percept is not a static value, the workspace should not necessarily broadcast:
label = "chair".
It may make available a dynamic structure:
M_chair(t)
or a compressed projection of it.
This means that globally available content may contain:
identity,
current state,
temporal context,
predictions,
affordances,
relevance.
11.13 A broadcast is not a copying of all neurons
Global availability does not mean that every module receives an exact copy of:
S(t).
Rather:
projection_i(M_A)
->
module_i.
For example, the motor system may make use of:
graspability.
The language module:
name.
The memory system:
similarity to previous experience.
All of them may, however, be influenced by the same underlying perceptual state.
11.14 The workspace and observables
For every module i we may define:
O_i = G_i(M_A).
The workspace therefore makes available different observables of the same dynamic object.
This is compatible with the conception of the Perceptual Manifold as a common internal reality available to different mechanisms through different projections.
11.15 Global availability and integration
If one state simultaneously influences:
memory,
action,
language,
value,
a functional integration arises.
The same percept is used across several subsystems.
This may be one of the main significances of global availability.
11.16 The workspace and the unity of behaviour
Without global access, different modules may respond to different local information.
The workspace may enable coordination:
current percept
->
consistent action,
consistent report,
consistent planning.
Global availability may thus contribute to the unified behaviour of the organism.
11.17 The workspace and attention
Attention may change the probability of:
workspace access.
For example:
attention(M_A)
->
amplification(M_A)
->
increased P(ignition).
Attention need not, however, be identical with the workspace.
It may function as a modulator of selection.
11.18 The workspace and prediction error
A strong unresolved prediction error may be one of the triggers of global access.
For example:
local model fails
->
prediction error persists
->
local state destabilizes
->
workspace ignition.
This may explain why unexpected events often gain wider processing.
11.19 The workspace and novelty
Similarly:
familiar predictable event
may be processed predominantly locally.
Conversely:
novel event
may have a high probability of:
global access.
The system therefore need not globally broadcast everything.
11.20 The workspace and relevance
A behaviorally significant state may be amplified even at low novelty.
For example:
threat,
reward,
goal relevance.
Therefore:
workspace priority
is not merely a function of prediction error.
11.21 The workspace and intuitive decision-making
An intuitive decision may arise before the whole process is globally available.
For example:
input
->
M_warning
->
action tendency.
The workspace may obtain only the resulting projection:
"something is wrong".
It need not obtain the whole history that led to:
M_warning.
This makes it possible to distinguish:
decision formation
from:
explicit reason representation.
11.22 Explicit reasoning
After workspace access the system may launch a further process:
M_A
->
global reasoning
->
new internal states.
Reasoning may in turn change:
attention,
prediction,
memory retrieval,
action selection.
A recurrent loop thereby arises between the workspace and the Perceptual Manifold.
11.23 The workspace is not merely an output buffer
If the workspace merely receives:
M_A
and sends it onward, that is too passive a model.
A global broadcast may in turn change the perceptual dynamics.
For example:
workspace content
->
attention shift
->
changed sensory gain
->
modified M.
Such a system is a closed dynamic loop.
11.24 The recurrent workspace
It may therefore be more appropriate to write:
perceptual manifold
<->
global workspace.
The workspace may:
read,
amplify,
modulate,
re-enter
the internal dynamics.
This is important for long-term conscious processing.
11.25 The workspace and working memory
A globally available state may be maintained longer thanks to:
recurrent workspace activity.
This need not, however, be the same as the original perceptual persistence.
It is important to separate:
percept persistence
and:
workspace maintenance.
11.26 A percept without the workspace
DPSH assumes the experimental possibility:
local M_A exists
but:
workspace disabled.
If the state still:
influences local behavior,
remains decodable,
persists,
then we have evidence for the separation of percept formation from global access.
11.27 A workspace without a stable percept
The opposite case is also possible.
The global system may be activated by:
noise,
error,
internal thought
without a stable externally anchored percept.
Therefore:
workspace activity
by itself need not be sufficient for perceptual content.
11.28 Workspace ablation
A key experiment:
perceptual network intact
workspace disabled.
We measure:
local state separability,
local context use,
cross-module accessibility,
explicit report.
DPSH expects the possibility of:
perceptual dynamics preserved
but:
global accessibility reduced.
11.29 Cross-module test
Let us imagine a percept:
M_A.
The visual subsystem is able to create the state.
The motor system is to respond according to it.
Without the workspace:
visual -> motor
need not be possible if the modules have no direct connection.
With the workspace:
visual M_A
->
workspace
->
motor.
Global availability can thereby be measured directly.
11.30 The workspace as a limited resource
Global Workspace Theory often works with the notion of a limited capacity.
DPSH may test this property by means of competition:
M_A
M_B
M_C.
If the workspace cannot globally stabilize all of them simultaneously, there arises:
access competition.
11.31 Workspace competition
Candidate states may compete in:
strength,
relevance,
novelty,
attention.
The result:
one or few states
->
global broadcast.
This may be a further symmetry-breaking process at a higher level.
11.32 Two levels of symmetry breaking
We may therefore distinguish:
Perceptual symmetry breaking
competing local interpretations
->
M_A.
Workspace symmetry breaking
competing perceptual states
->
global access(M_A).
These processes need not be identical.
11.33 The workspace as a further dynamic manifold
The workspace itself may have its own state space:
W(t).
The global dynamics is then not:
M -> output,
but:
M(t)
<->
W(t).
The system may have two mutually interacting dynamic structures:
perceptual manifold P
workspace manifold W.
11.34 The coupling of P and W
Formally:
dP/dt = F(P, W, I)
and:
dW/dt = G(W, P, goals, memory).
Such an architecture allows:
bottom-up access
and:
top-down modulation.
11.35 Ignition threshold
For the workspace we may define a macroscopic quantity:
G(t)
for example the degree of global recurrent activation.
Ignition occurs at:
G(t) > θ_G.
It is important to test whether there is a genuine nonlinear transition, or merely a smooth increase of availability.
11.36 Ignition and stochasticity
If a candidate state is just below the threshold:
G ≈ θ_G,
a small stochastic fluctuation may cause:
ignition.
This may lead to trial-to-trial variability with an identical near-threshold stimulus.
11.37 Ignition and phase
Workspace access may similarly depend on the temporal configuration.
The same perceptual input:
M_A
may gain global access in a certain phase:
φ_good
and fail at:
φ_bad.
In this way the connection of local temporal organization with the global broadcast can be tested.
11.38 Ignition and history
If the workspace itself has a state:
W(t),
then its response to:
M_A
depends on:
W_previous.
Global access may therefore exhibit:
hysteresis,
refractory effects,
history dependence.
11.39 Explicit report
A report can be understood as a downstream function:
workspace content
->
language/report module.
If:
report = A,
this does not automatically mean that the report creates the percept.
It is merely one of the observables of the globally available state.
11.40 Report-free measurement
For research it is important also to use metrics independent of the explicit report.
For example:
forced-choice action,
prediction,
state decoding,
perturbation response.
In this way the perceptual dynamics can be separated from the report mechanism itself.
11.41 The workspace and the phenomenal hypothesis
One possibility is:
phenomenal experience
depends on
global access.
Another:
phenomenal experience
depends on
perceptual state before access.
DPSH does not decide this question at this stage.
It is precisely the separation of:
percept formation
and:
global availability
that makes it possible to test these alternatives later.
11.42 Possible models of the relation between percept and consciousness
Model A
percept
->
workspace
->
consciousness.
Model B
percept = phenomenal state
and the workspace merely provides:
access/report.
Model C
recurrent interaction
between percept and workspace
is necessary for phenomenal experience.
DPSH must remain compatible with the testing of all three possibilities.
11.43 Experiment GW1 – local percept without the workspace
We create the task:
A -> blank -> ambiguous X.
The network creates:
M_A.
The workspace is subsequently switched off.
We test:
state persistence,
local decision,
explicit report,
cross-module transfer.
If the local functions remain but global access disappears, this supports the separation of the two mechanisms.
11.44 Experiment GW2 – workspace ablation after percept formation
We first let the following arise:
M_A.
Only then do we carry out:
workspace ablation.
We track whether:
M_A persists
and whether the only change is in:
global accessibility.
11.45 Experiment GW3 – workspace ablation before percept formation
Conversely, we switch off the workspace before the stimulating input.
If:
M_A
does not arise at all, this may mean that the workspace is necessary for percept formation.
We thereby obtain a direct test of the relation.
11.46 Experiment GW4 – cross-module accessibility
The percept arises only in the visual subsystem.
We then test whether the information can be used by:
motor,
memory,
language.
We compare:
workspace ON
versus:
workspace OFF.
11.47 Experiment GW5 – ignition threshold
We progressively vary the strength of the stimulus:
λ.
We measure:
local percept strength
and:
global workspace activity.
We look for whether the workspace exhibits:
smooth response
or:
nonlinear threshold / ignition.
11.48 Experiment GW6 – competition
We simultaneously create:
M_A
M_B.
Both compete for the workspace.
We vary:
relevance,
sensory strength,
attention.
We track:
which state gains access.
11.49 Experiment GW7 – novelty
We compare:
familiar predictable stimulus
and:
novel stimulus.
We measure:
workspace access probability,
prediction error,
local percept formation.
The hypothesis assumes that novelty may influence access even at a similar perceptual strength.
11.50 Experiment GW8 – unresolved prediction error
The network receives an input that the local perceptual model is unable to explain well.
We measure:
persistent prediction error
->
workspace activation.
In this way we test the hypothesis that the workspace resolves above all locally unresolved situations.
11.51 Experiment GW9 – intuitive action before report
The network learns a complex classification situation.
After a new input we measure the time:
t_action
and:
t_workspace/report.
If:
t_action < t_report,
the decision may arise before explicit global availability.
11.52 Experiment GW10 – top-down feedback
After workspace ignition we change the top-down signal:
attention to feature X.
We measure the change in:
local perceptual dynamics.
If the workspace genuinely forms a recurrent loop, the global state must influence the Perceptual Manifold in return.
11.53 Experiment GW11 – workspace phase dependence
With the same local percept we vary the:
phase relationship
between the perceptual network and the workspace.
We measure:
ignition probability,
transmission efficacy,
global availability.
This connects the Global Workspace with the oscillatory part of DPSH.
11.54 Experiment GW12 – capacity limit
We present several simultaneously relevant percepts:
M_1 ... M_n.
We progressively increase:
n.
We measure:
number globally accessible,
interference,
switching.
In this way the limited capacity of the workspace can be tested.
11.55 A metric of global accessibility
We define:
A_global(M)
as the number or range of functionally distinct modules whose behaviour
can be causally influenced by the state M.
A low value:
local state.
A high one:
globally accessible state.
11.56 A metric of ignition
We may measure:
G_peak,
propagation range,
recurrence duration,
number of activated modules.
It will be important to distinguish a genuine nonlinear ignition from a mere increase of overall activity.
11.57 A metric of workspace selectivity
For the candidates:
M_A,
M_B
we define:
P(access_A),
P(access_B).
Manipulation of relevance or attention should systematically change their probability.
11.58 Falsification criteria
The proposed division between DPSH and the Global Workspace will be weakened if:
- no perceptual state arises without an active workspace,
- workspace ablation always destroys the local perceptual representation as well,
- global availability cannot be experimentally separated from the existence of a perceptual state,
- no measurable cross-module broadcast occurs,
- the workspace exhibits no competition or selectivity,
- the assumed ignition is fully explicable by a linear increase of activity,
- top-down workspace feedback has no influence on the perceptual dynamics,
- global availability brings no functional capacity beyond local processing.
In such a case it will be necessary to reformulate fundamentally the relation between DPSH and the Global Workspace.
11.59 Chapter research hypothesis
We formulate the partial hypothesis H10:
H10 – Global Accessibility Hypothesis
A metastable perceptual state and its global availability are functionally distinct processes. Perceptual dynamics may arise in a distributed recurrent network before global access, while the Global Workspace enables a selected dynamic state to gain causal influence over a broad spectrum of otherwise specialized subsystems. Access to the workspace may arise through a nonlinear recurrent ignition process modulated by relevance, attention, prediction error and the state of the system.
The stronger falsifiable prediction:
If percept formation and global accessibility genuinely constitute separable mechanisms, there must exist an experimental condition in which a decodable and causally relevant local perceptual state can be preserved while its availability to distant modules, to explicit report or to the wider decision-making system is reduced.
DPSH thereby proposes:
percept formation
!=
global accessibility.
And it admits at the same time that:
phenomenal consciousness
may depend on one of these mechanisms or on their recurrent interaction.
This question cannot be decided by the function of the Global Workspace alone and will be the subject of the phenomenal part of the hypothesis.
12. The Perceptual Manifold as a shared internal model of the world
12.1 From the individual percept to the inner world
The preceding chapters worked predominantly with individual metastable perceptual states:
M_A,
M_B,
M_C.
This is still not sufficient, however, to describe continuous experience.
An organism does not perceive only isolated objects or properties.
At one moment there may simultaneously exist a dynamic representation of:
space,
objects,
motion,
one's own body,
sounds,
expectations,
relevance,
possible actions.
The Dynamic Perceptual State Hypothesis therefore introduces a broader object:
Perceptual Manifold.
This constitutes a continuously existing structure of the internal state space in which the individual perceptual states are mutually interconnected.
12.2 The Perceptual Manifold is not an image
The Perceptual Manifold is not understood as an internal bitmap or an exact copy of the sensory input.
It is not an:
internal screen.
It is a dynamic representation of relations.
An object, for example, may be represented simultaneously through:
identity,
position,
motion,
relation to body,
behavioral relevance,
predicted continuation.
It is therefore more appropriate to understand the manifold as a:
structured dynamic state
rather than as a:
reconstructed image.
12.3 The internal model is not identical with the external world
The Perceptual Manifold is not the world.
It is an internal dynamic construction of the system.
Formally:
World(t)
->
sensory channels
->
neural dynamics
->
P(t),
where:
P(t) != World(t).
P(t) is merely the state that enables the system to predict, interpret
and influence the world effectively.
12.4 Functional adequacy instead of complete reconstruction
The internal model need not contain all the physical properties of the environment.
It suffices if it contains the structure relevant for:
prediction,
action,
memory,
valuation,
planning.
DPSH therefore does not assume:
perfect world reconstruction.
It assumes:
functionally adequate internal model.
12.5 The dynamic structure
The Perceptual Manifold can be provisionally represented as:
P(t) =
{
active perceptual regions,
latent states,
phase relations,
transition probabilities,
predictions,
learned constraints
}.
This object is not explicitly stored in one place.
It is the result of the joint dynamics of the entire relevant network.
12.6 A shared internal state
The key hypothesis of this chapter is:
Several functional subsystems may work over different projections of the same internal dynamic state, instead of each module creating its own independent reconstruction of the world.
Schematically:
sensory systems
|
v
Perceptual Manifold
|
+-----+-----+------+-------+-------+
| | | | |
v v v v v
memory action value language planning.
12.7 Projection instead of a copy
Every module does not need the whole state:
P(t).
It may obtain only the functionally relevant projection:
O_i(t) = G_i(P(t)).
For example:
Motor module
O_motor =
{
reachable,
direction,
obstacle,
movement opportunity
}.
Language module
O_language =
{
object identity,
relational concepts,
symbolic labels
}.
Valuation module
O_value =
{
threat,
reward,
salience
}.
Different modules therefore read different properties of the same internal dynamics.
12.8 One representation, many observables
An analogy thereby arises:
one underlying state
->
multiple observables.
The same internal state may be interpreted differently from the point of view of different modules.
For example, an object:
apple
may simultaneously be:
red,
edible,
reachable,
familiar,
named "apple".
There need not exist five independent worlds.
There may exist one distributed structure with several functional projections.
12.9 The Perceptual Manifold as a frame of reference
If several subsystems use the same internal state, they may share a reference.
For example:
visual system:
object A is left of object B
motor system:
move hand toward A
language system:
"the object on the left"
All these operations may be anchored in the same relational structure.
12.10 Relational representation
DPSH therefore assumes that relations are more important than isolated properties:
A left_of B,
A moving_toward B,
A occludes B,
hand near A,
A predicts B.
The Perceptual Manifold may be rich precisely in that it retains a network of these dynamic relations.
12.11 Space as a relation
A spatial representation need not necessarily be an explicit Cartesian map.
It may arise from relational structures:
near,
far,
left,
right,
above,
behind,
reachable.
A global spatial percept may be the emergent geometry of these relations.
12.12 Time as a relation
Likewise, time need not be stored only as a:
timestamp.
It may be represented as:
before,
after,
expected next,
duration,
phase relation.
The Perceptual Manifold is therefore a spatiotemporal dynamic object.
12.13 An object as a stable trajectory of relations
Object identity may be represented by the continuity of dynamic relations.
An object, for example, moves:
x1 -> x2 -> x3.
Its sensory pattern changes.
Nevertheless:
identity_relation persists.
An object can be understood as a stable structure inside a changing manifold.
12.14 One's own body as part of the manifold
DPSH does not assume a sharp separation of:
world representation
and:
body representation.
The internal state may include:
body position,
proprioception,
action capability,
interoceptive state.
This is important, because action changes future sensory input.
12.15 The egocentric frame of reference
Part of the inner world may be organized with respect to the organism:
left of me,
reachable by me,
approaching me.
Such a representation is not a purely objective map.
It is functionally related to the current possibilities of the system.
12.16 Affordances
An object need not be represented only by the question:
"what is it?"
It may at the same time contain:
"what can be done with it?"
For example:
cup
->
graspable,
drinkable,
movable.
An affordance may be a projection of the same perceptual state into the motor system.
12.17 Value as a modulation of the manifold
The valuation system may change the dynamic geometry:
threat
->
increased stability / salience.
An object with high relevance may be:
easier to enter,
harder to ignore,
more likely to gain workspace access.
Value therefore need not be merely a tag attached to a finished percept.
It may directly deform its dynamics.
12.18 Emotional and interoceptive state
The internal state of the body may similarly influence the interpretation of the same environment.
For example:
hunger
may change the stability of the representations:
food-related states.
The Perceptual Manifold is therefore potentially dependent not only on external sensors, but also on internal states of the organism.
12.19 Context
The same object:
X
may be interpreted differently in:
context A
and:
context B.
Formally:
M_X(context A)
!=
M_X(context B).
Context is not external metadata.
It is part of the present dynamic state.
12.20 Contextual deformation of the manifold
Context may change the:
basin geometry,
transition probabilities,
prediction priors.
Some interpretations thereby become:
more likely
and others:
less likely.
12.21 A hierarchy of manifolds
It is possible that there does not exist only one homogeneous manifold.
There may exist a hierarchy:
local sensory manifolds
->
regional integrative manifolds
->
global perceptual manifold.
For example:
visual manifold
auditory manifold
body manifold
may connect into a broader:
multimodal state.
12.22 Multimodal integration
The same object may generate:
visual input,
sound,
touch.
If all of them correspond to the same external cause, the internal dynamics may create an integrated state:
M_object.
This is not a simple sum of:
visual + audio + touch.
It may contain their common relational structure.
12.23 Cross-modal prediction
A visual state may predict a sound:
falling object
->
expected impact sound.
A sound may conversely change visual expectations:
sound behind
->
orient visual system.
The Perceptual Manifold thus enables predictions across modalities.
12.24 Binding as common dynamic compatibility
Feature binding need not be resolved by a separate:
binder.
If:
color,
position,
motion,
shape
mutually support the same dynamic configuration, they may become part of:
M_object.
Binding then arises as a stable compatibility within the manifold.
12.25 Incompatible bindings
If two properties cannot be stabilized simultaneously in one configuration, the network may create a competition:
M_A
M_B.
Symmetry breaking subsequently selects one interpretation.
Binding is thereby connected with the preceding mechanisms.
12.26 The Perceptual Manifold as a predictive object
The Perceptual Manifold does not describe only:
what is.
It contains the dynamics of:
what can happen next.
Therefore:
P(t)
implicitly contains a:
transition model.
12.27 The inner world as a generative structure
From the present state:
P(t)
it is possible to generate expectations:
predicted sensory stream,
predicted object motion,
predicted consequences of action.
The inner world is therefore an active generative model.
12.28 Counterfactual dynamics
A stronger architecture may make it possible to simulate internally:
what if action A?
without the immediate realization of the action.
For example:
P(t)
->
simulated transition A
->
expected state P_A'.
This would enable planning.
12.29 Simulation is not necessary for the basic percept
DPSH does not, however, require counterfactual simulation for the basic formation of a perceptual state.
It is a higher function that may be built over the manifold later.
12.30 Action changes the internal model
An action:
a(t)
changes the world:
World(t)
->
World(t + dt).
This changes the future sensory input:
I(t + dt).
The Perceptual Manifold is therefore part of a closed loop:
perception
->
action
->
world
->
new perception.
12.31 Active perception
The system need not merely wait passively for data.
It may carry out actions that reduce uncertainty:
move eyes,
turn head,
approach object.
It thereby actively obtains the information needed to stabilize the internal model.
12.32 The Perceptual Manifold and decision-making
Decision-making can be understood as the selection of an action according to the present dynamic state:
action =
F(P(t), goals, value).
The action system need not receive the whole sensory stream.
It may work with a projection of the manifold.
12.33 Intuitive decision-making
If experience has created a certain geometry:
P_learned,
a new complex input may rapidly move the system into:
M_action_A.
The action may be selected without an explicit computation of all the reasons.
Intuition is thereby naturally connected with the shared internal model.
12.34 Explicit reasoning
Reasoning may carry out slower transformations over the manifold.
For example:
current state
->
retrieve memory
->
simulate alternative
->
modify prediction
->
new state.
Reasoning therefore need not be the source of the basic percept.
It may be a mechanism for manipulating an already existing internal model.
12.35 Language as a projection of the inner world
The language system may convert parts of the dynamic state into symbols:
P(t)
->
linguistic representation.
It thereby gains the possibility of reporting the content of the percept.
12.36 A symbol is not a percept
The word:
"chair"
is not identical with:
M_chair.
It is merely one projection of this state.
This explains why an internal representation may contain markedly more information than a verbal report.
12.37 Memory as a link to a past manifold
Memory may retain:
previous states,
compressed trajectories,
synaptic traces.
Upon recall:
memory
->
current manifold.
A recollection may thus reactivate parts of an earlier dynamic structure.
12.38 A recollection is not necessarily an exact reconstruction
Because the present state:
P_now
differs from the state at the time of the original experience, the reactivation may be:
reconstructive.
A recollection may be influenced by the current context.
12.39 The Perceptual Manifold as the working space of experience
From a functional point of view, the Perceptual Manifold can be understood as the state in which the following meet:
sensory evidence,
memory,
prediction,
valuation,
action possibilities.
This does not mean that it is one anatomical region.
It is a distributed dynamic structure.
12.40 The difference from the Global Workspace
The Perceptual Manifold and the Global Workspace are not the same.
The Perceptual Manifold:
represents and evolves internal world state.
The Global Workspace:
provides broader accessibility to selected content.
The working relation:
Perceptual Manifold
->
selected state
->
Global Workspace
->
distributed access.
12.41 The workspace need not transmit the whole manifold
A global broadcast may contain only the relevant part:
projection(P).
For example:
"unexpected moving object on left".
The whole internal spatial model need not be globally broadcast.
12.42 The workspace may change the manifold in return
After global access:
workspace
->
attention,
memory retrieval,
action,
prediction.
These processes influence in return:
P(t).
There arises:
P <-> W.
12.43 The Perceptual Manifold and the continuity of the subject
A stronger, so far theoretical possibility is that the continuity of subjective experience is related to the fact that:
P(t)
never begins from a zero state.
Every new moment arises by a transformation of:
P(t - dt).
An uninterrupted causal trajectory of the inner world thereby arises.
12.44 "The same world" as dynamic invariance
Even though:
P(t1) != P(t2),
a higher invariance may exist:
same scene,
same objects,
same self.
The continuity of experience may therefore exist at the macroscopic level despite constant microscopic change.
12.45 The internal perspective
The Perceptual Manifold is not a neutral physical map.
It is organized from the perspective of the system.
It contains relations such as:
relative to body,
relevant to goals,
reachable,
dangerous,
expected.
It is therefore a model of:
world-for-the-system
rather than:
world-in-itself.
12.46 Identity of the representation across modules
If several modules work over the same state, it is possible to test whether their different outputs genuinely correlate with a single latent structure.
For example, the state:
M_A
should simultaneously predict:
motor choice,
verbal label,
memory retrieval,
value response.
This provides an experimental test of the shared internal model.
12.47 Experiment PM1 – common latent state
The network receives a complex scene.
We record:
global population state.
Then several modules perform different tasks:
classify object,
choose action,
estimate value,
predict next state.
We test whether their results can be explained by projections of the same latent state.
12.48 Experiment PM2 – shared-state perturbation
We perturb the region:
M_A.
If it is genuinely the shared basis of several functions, the perturbation should simultaneously change several downstream outputs:
action,
report,
prediction.
If it changes only one isolated module, this may rather be a local representation.
12.49 Experiment PM3 – cross-modal completion
The network learns an object by means of:
vision + sound.
Later it receives only:
partial visual input.
We test whether the internal state generates the expectation of:
corresponding sound.
This would support the existence of an integrated multimodal representation.
12.50 Experiment PM4 – occluded object
An object moves and is subsequently occluded.
The Perceptual Manifold is to maintain:
identity,
predicted position,
expected reappearance.
We measure whether this information can be used by:
motor,
prediction,
report
even without the current visual input.
12.51 Experiment PM5 – conflicting modalities
The visual and the auditory channel receive incompatible information.
We track:
competition,
state formation,
confidence,
workspace access.
In this way it is possible to test how the manifold resolves a multimodal conflict.
12.52 Experiment PM6 – module removal
We remove one downstream module:
language.
If the Perceptual Manifold exists independently, the following should remain:
perception,
action,
prediction.
This will help to separate the internal state from its particular projections.
12.53 Experiment PM7 – sensory modality removal
After multimodal learning we remove:
vision
or:
audio.
We track whether the remaining modality is able to activate part of the common manifold.
12.54 Experiment PM8 – context-dependent interpretation
We present the same stimulus:
X
in:
context A
context B.
We measure:
M_XA
M_XB.
If the manifold genuinely integrates context, the states should differ even with the same local input.
12.55 Experiment PM9 – action-dependent perception
The network carries out an action:
move sensor.
It thereby obtains a new input.
We test whether the update:
P(t)
->
action
->
I(t + dt)
->
P(t + dt)
preserves object and spatial continuity.
12.56 Experiment PM10 – counterfactual planning
A later extension:
simulate action A
simulate action B.
Without an actual action we track the predicted:
P_A'
P_B'.
If these states can be used to select an action, the manifold also functions as a basis for generative planning.
12.57 Experiment PM11 – shared latent decoder
We train several simple decoders over the same population state:
decoder_object,
decoder_position,
decoder_value,
decoder_action.
If all of them use a common latent structure, they should be able to obtain the relevant information without a separate reconstruction of the sensory data.
12.58 Experiment PM12 – causal projection test
We change one macroscopic property of the manifold, for example:
object position.
We track whether the following change consistently:
motor target,
spatial language,
predicted movement.
This tests whether the modules genuinely share the same reference structure.
12.59 A metric of integration
For a set of modules:
G1 ... Gn
we may measure how many of their outputs can be predicted from the common state:
P(t).
Higher shared predictive information may indicate a shared latent representation.
12.60 A metric of cross-modal consistency
For an object represented by several modalities we measure:
consistency(
visual projection,
auditory projection,
action projection
).
An integrated manifold should create mutually compatible projections.
12.61 A metric of continuity under sensor transformation
We change the sensory input markedly, but preserve the object.
For example:
illumination,
rotation,
occlusion.
We measure whether the state remains in the:
same macrostate family.
In this way we test dynamic invariance.
12.62 A metric of shared causal impact
We perturb the latent state.
We measure the change in several modules:
Δaction,
Δprediction,
Δreport,
Δvaluation.
If one perturbation consistently influences several functions, this supports the hypothesis of a shared internal model.
12.63 Falsification criteria
The hypothesis of a shared Perceptual Manifold will be weakened if:
- the individual modules need completely independent representations of the same environment,
- a perturbation of the perceptual state does not influence several downstream functions,
- multimodal information cannot be integrated into common states,
- object continuity disappears upon a change of modality or a brief occlusion,
- the same latent states do not support different functional projections,
- every new task requires a new reconstruction of the original sensory data,
- the common dynamic state provides no advantage over a set of independent modular representations,
- the internal manifold cannot be causally distinguished from a mere analytical construction of the observer.
In such a case the notion of the Perceptual Manifold would have to be restricted to a descriptive state-space analysis instead of a functional shared internal model.
12.64 Chapter research hypothesis
We formulate the partial hypothesis H11:
H11 – Shared Perceptual Manifold Hypothesis
Continuous perception may be realized through a distributed dynamic state space that simultaneously represents the mutually dependent properties of the environment, the body, the context and the expected development. Different functional subsystems need not create their own independent reconstructions of the world, but may obtain different projections of this common internal state.
The stronger falsifiable prediction:
If the Perceptual Manifold genuinely functions as a shared internal model, then a targeted change of its macroscopic state must cause consistent changes in several functionally distinct subsystems, while the removal of an individual downstream module must not by itself destroy the basic perceptual representation.
DPSH thereby proposes:
sensory streams
->
shared dynamic internal world
->
multiple functional projections.
Not:
sensory streams
->
separate reconstruction for every task.
This chapter constitutes the transition from the individual percept to the broader hypothesis of a continuous inner world.
Only over such a state can the question then be posed precisely as to whether some of its properties may be related to phenomenal experience.
13. Phenomenal state and qualia
13.1 The hardest part of the hypothesis
The preceding chapters formulated a mechanistic model that is, at least in principle, experimentally testable.
They described the possibility that a globally clockless recurrent network may, through:
stochasticity,
local oscillations,
non-commutative timing,
recurrent amplification,
symmetry breaking,
metastability,
hysteresis,
predictive constraints,
deep state learning
create an ongoing internal dynamic model of the world.
Subjective experience is thereby still not explained.
The question remains:
Why should a certain dynamic state of a system at the same time be a state that is somehow experienced?
This question constitutes the transition from the functional problem of perception to the phenomenal problem of consciousness.
13.2 A percept is not automatically a quale
DPSH consistently distinguishes:
perceptual state
and:
phenomenal state.
A perceptual state may be:
decodable,
persistent,
predictive,
causally relevant,
globally available,
without our having thereby demonstrated that the system subjectively experiences anything.
Therefore the following does not hold:
functional percept
=>
phenomenal experience.
This logical step must remain explicitly open.
13.3 A working definition of a quale
For the purposes of DPSH we use the notion quale to designate the phenomenal quality of experience.
For example:
what it is like to see red,
what it is like to feel pain,
what it is like to hear a certain tone.
A quale is therefore not the information itself:
wavelength = X
nor:
classification = red.
It designates the subjective character of the given percept.
13.4 Phenomenal content versus report
It is also necessary to separate:
phenomenal content
from:
report about phenomenal content.
A system may create the output:
"I see red"
on the basis of a learned symbolic mapping.
Such a report does not by itself demonstrate:
experience of red.
A report therefore cannot be the main criterion of level E3.
13.5 The strong phenomenal hypothesis of DPSH
DPSH formulates the following stronger hypothesis:
Phenomenal content may be associated not with the activation of an individual neuron, symbol or static representing pattern, but with the particular organization of a temporally extended metastable dynamic state of the whole relevant neuronal system.
Schematically:
temporally organized
metastable population state
?
->
phenomenal content.
The question mark is part of the hypothesis.
This relation has not so far been demonstrated.
13.6 Why a dynamic state?
The reason for this hypothesis is not the claim that dynamics itself creates consciousness.
The motivation lies in several properties of phenomenal experience:
continuity,
integration,
context dependence,
temporal duration,
unity,
the capacity to change without a loss of percept identity.
These properties have a natural functional counterpart in a metastable dynamic state.
13.7 Phenomenal identity as macroscopic invariance
If the same quale can persist despite the constant exchange of individual spikes, then its neuronal correlate probably need not be an exact microscopic pattern.
A working possibility:
many microscopic realizations
->
same macroscopic phenomenal state.
That is:
S1 != S2
but:
S1 in M_red
S2 in M_red.
Phenomenal identity could then be a property of the macrostate.
13.8 A quale as a dynamic region, not a point
The stronger version assumes:
quale_A ~ M_A
where M_A is not a single instantaneous state.
It is a region of the dynamics in which the system may exhibit:
internal movement,
stochastic variability,
phase evolution,
and nevertheless remain within the same phenomenal content.
13.9 The temporal dimension of qualia
Phenomenal experience is evidently not an event of zero duration.
The percept:
red
exists for a certain time.
A more appropriate model may therefore be:
quale(t0, t1)
rather than:
quale(t).
DPSH therefore admits that phenomenal content may be a property of a temporally extended trajectory.
13.10 The minimal phenomenal interval
If a quale is a dynamic process, there may be a minimal temporal window:
Δt_q
needed for its formation.
Too brief an activation might create:
local neural response
without:
stable phenomenal state.
This is an experimentally interesting but hard-to-verify prediction.
13.11 Oscillatory organization and phenomenal content
If relative phase genuinely contributes to the identity of a perceptual state, the following may also be phenomenally relevant:
phase geometry.
The strong hypothesis would therefore not be:
firing rate = quale,
but rather:
population state
+
temporal phase organization
->
candidate phenomenal substrate.
13.12 Stochasticity and phenomenal variability
Microscopic stochasticity means that two realizations of the same percept need not be identical.
Nevertheless the subjective content may remain stable.
This supports the working idea that a phenomenal state must be invariant with respect to part of the microscopic noise.
13.13 Non-commutativity and experience
If the order of events changes the internal state:
A -> B
!=
B -> A,
it may also change the phenomenal context of the subsequent percept.
The same stimulus X may be experienced differently depending on:
previous state,
expectation,
emotional context,
attention.
This is compatible with history-dependent dynamics.
13.14 Phenomenal context
A quale need not exist in isolation.
The colour of an object, for example, is experienced in the context of:
the surrounding colours,
the light,
object identity,
attention.
It is therefore possible that:
quale_A
is not a property of an isolated subsystem, but a local property of the broader Perceptual Manifold.
13.15 The relationality of phenomenal content
DPSH therefore admits:
Phenomenal quality may be determined not only by the state of one representation, but by its relations to the other active and latent states of the system.
That is:
Q_A =
F(
M_A,
context,
phase relations,
body state,
predictions
).
This is stronger than the notion of a single "qualia neuron".
13.16 Phenomenal space
If many phenomenal states exist:
Q1,
Q2,
...,
Qn,
they may form their own structure of similarities and transitions.
Some colours, for example, are phenomenally closer to each other than others.
One may thereby consider:
phenomenal state space.
DPSH may later ask whether its geometry correlates with the geometry of the Perceptual Manifold.
13.17 Structural correspondence
One of the strongest empirically accessible routes may be:
neural manifold geometry
<->
phenomenal similarity geometry.
For example, if a subject judges:
Q_A similar to Q_B
Q_A dissimilar to Q_C,
then it is possible to test whether:
D(M_A, M_B)
<
D(M_A, M_C).
Such a result still does not demonstrate an identity between the two spaces.
It would, however, show a structural correspondence.
13.18 Phenomenal structure instead of phenomenal substance
DPSH need not claim that a quale is some new physical substance.
The stronger working direction is:
Phenomenal properties may correspond to a specific organization and to relations within a dynamic physical system.
That is:
no extra substance required
but:
special organization may matter.
This is the metaphysical minimum of the hypothesis.
13.19 Supervenience as a working assumption
DPSH may work with the assumption:
A change of the phenomenal state requires a change of the relevant physical dynamic state.
That is, if:
Q_A != Q_B,
then there must exist some difference:
M_A != M_B.
This does not resolve why a physical state has phenomenality.
But it provides a testable relation between changes.
13.20 Identity is not demonstrated
Even a perfect correlation:
Q_A <-> M_A
does not demonstrate:
Q_A = M_A.
It may be merely a:
neural correlate.
DPSH must therefore consistently distinguish:
correlation,
causal necessity,
sufficiency,
identity.
13.21 Four levels of evidence
For every candidate mechanism we distinguish:
Level 1 – Correlation
M_A occurs when Q_A is reported.
Level 2 – Causal relevance
A perturbation of M_A changes the phenomenal report or the
discrimination.
Level 3 – Necessity
Without M_A, Q_A does not occur.
Level 4 – Sufficiency
An artificial creation of M_A evokes Q_A.
Even level 4, however, does not by itself resolve the philosophical question of identity.
13.22 The problem of an artificial network
With a biological subject it is at least possible to obtain:
report,
behavior,
neural measurements.
With Cognia there is no independent access to subjective experience.
If Cognia says:
"I am experiencing red",
this may be a learned output.
It is therefore not possible to demonstrate artificial phenomenal experience from a report alone.
13.23 What can genuinely be tested in Cognia
Cognia can test the mechanisms that the hypothesis regards as a candidate substrate:
metastability,
temporal integration,
shared manifold,
history dependence,
phase sensitivity,
predictive persistence,
global accessibility.
This will make it possible to test:
functional architecture of deep percept.
Not directly:
existence of qualia.
13.24 Deep percept
To separate the two levels we introduce the working term:
Deep Percept.
Deep Percept designates an internal dynamic state that:
- is not merely an immediate reaction to a sensor,
- contains history,
- persists across changes of input,
- is distributed,
- is multimodal or integrative,
- generates predictions,
- influences several downstream systems,
- has its own metastable dynamics.
Deep Percept is functionally testable.
A quale is not, directly.
13.25 Deep Percept versus quale
The working relation:
sensory processing
->
deep percept
->
?
phenomenal quale.
The whole research programme of DPSH may therefore first test the formation of:
Deep Percept.
Only if this mechanism works is there any point in discussing its phenomenal interpretation.
13.26 Phenomenal unity
One of the important properties of conscious experience is its apparent unity.
At a single moment we do not experience independently:
color quale,
shape quale,
sound quale,
body quale
as perfectly isolated systems.
They are part of one broader state of experience.
DPSH seeks the functional counterpart of this unity in an integrated Perceptual Manifold.
13.27 Unity does not mean homogeneity
A global phenomenal state need not be one uniform activation.
It may contain:
many local processes
connected into a:
coherent dynamic whole.
Unity may therefore be relational.
13.28 The Global Workspace and phenomenality
There are at least three possible relations.
Model A – access consciousness
Phenomenal experience arises only at:
Deep Percept
->
Global Workspace.
Model B – perceptual phenomenality
A Deep Percept may be phenomenal already before global access.
The workspace merely enables:
report,
reasoning,
broad control.
Model C – recurrent unity
A phenomenal state requires a closed interaction:
Perceptual Manifold
<->
Global Workspace.
DPSH does not so far decide among these possibilities.
13.29 Phenomenality and the Global Workspace report
If an experiment shows:
perceptual state survives workspace ablation
but:
report disappears,
this demonstrates only the separation of:
percept
and:
report/global access.
It does not automatically say whether phenomenal experience persisted.
13.30 Phenomenal continuity
If the phenomenal state is associated with the Perceptual Manifold, its continuity may arise from:
P(t + dt) = F(P(t), I(t)).
Every new state has a causal continuity with the previous one.
This provides a candidate mechanism for the continuous stream of experience.
13.31 The stream of consciousness
The working model:
Q(t0)
->
Q(t1)
->
Q(t2)
-> ...
need not be a sequence of separate snapshots.
It may be a single continuous dynamic trajectory:
Q(t).
If DPSH is right, the neural counterpart may be:
P(t).
13.32 Phenomenal transitions
A change of percept need not be instantaneous.
For example:
M_A
->
transition region
->
M_B.
It is possible that the subjective change:
Q_A -> Q_B
corresponds precisely to this macroscopic change of the dynamic regime.
13.33 Perceptual rivalry as a test window
Ambivalent perception is especially interesting.
The sensory input remains:
constant.
The phenomenal report, however, passes:
Q_A
->
Q_B.
If the neural dynamics passes:
M_A
->
M_B
without a change of input, we obtain a stronger connection between the internal dynamic state and the conscious percept than in a mere stimulus-response correlation.
13.34 Phenomenal hysteresis
If the reported percept exhibits hysteresis:
Q_A persists
despite a certain change of input,
and the neural state exhibits the same:
M_A persistence,
it is possible to test a structural link:
neural hysteresis
<->
phenomenal hysteresis.
13.35 Phase manipulations
A very important possible test follows from the preceding chapters.
If:
phase scramble
changes:
M_A
with an approximately preserved firing rate,
and at the same time changes the subjective percept,
then relative timing is a candidate component of phenomenally relevant dynamics.
This type of experiment is stronger than a correlation of firing rate with report.
13.36 Timing without a change of content
The opposite result is equally important.
If it is possible to change markedly the:
temporal organization
without a change of:
reported quale,
then the strong hypothesis that phase geometry determines phenomenal identity will be weakened.
13.37 The macroscopic sufficiency hypothesis
A stronger variant of DPSH might claim:
If two physical realizations share a relevantly identical macroscopic dynamic organization, they should correspond to the same phenomenal content regardless of microscopic differences.
Schematically:
macrostate(M1) ≈ macrostate(M2)
hence:
Q1 ≈ Q2.
This is a very strong and so far undemonstrated hypothesis.
13.38 Substrate independence
From the preceding claim there would potentially follow:
same relevant dynamics
->
same phenomenal organization
even on a different physical substrate.
This would make artificial phenomenal experience possible.
DPSH does not, however, accept this conclusion as fact so far.
It is a further degree of the hypothesis.
13.39 Biological specificity as an alternative
It is possible that the correct dynamic architecture is not itself sufficient.
Phenomenality may require:
specific biology,
cellular mechanisms,
chemical modulation,
unknown physical processes.
Cognia may then create:
perfect functional Deep Percept
without:
phenomenal experience.
This alternative cannot be excluded in advance by experiments within Cognia itself.
13.40 Functional isomorphism
If Cognia reproduces:
state dynamics,
prediction,
integration,
hysteresis,
global access,
adaptive learning,
we may claim:
functional similarity.
We cannot without more claim:
phenomenal equivalence.
13.41 The epistemic limit
A fundamental problem is that phenomenal experience is directly accessible only to the subject who experiences it.
From the outside we have access only to:
behavior,
report,
physical dynamics.
No theory of consciousness can therefore simply measure a quale in the same way as a firing rate.
13.42 Inference to the best explanation
The phenomenal hypothesis can therefore be supported only indirectly.
For example, if a certain dynamic model simultaneously explains:
perceptual unity,
bistability,
temporal continuity,
context dependence,
report,
action,
neural perturbation effects,
it may constitute a better explanation than simpler alternatives.
This is, however, still an inference and not a direct observation of qualia.
13.43 Minimal phenomenal conditions
DPSH may later attempt to identify a minimal functional set:
M = {
recurrent dynamics,
temporal depth,
integration,
history dependence,
self-maintenance,
global or broad accessibility
}.
One may then ask:
Which of these properties are necessary for a conscious percept in a biological system?
This is empirically more accessible than the question:
"what is a quale in itself?"
13.44 The perturbational approach
The strongest experiments will not merely track correlations.
They will actively disrupt the candidate mechanisms:
phase,
recurrence,
metastability,
workspace coupling.
And track:
changes in subjective perception.
In this way the causally necessary mechanisms can be identified.
13.45 Experiment Q1 – neural/phenomenal geometry
With biological data:
- obtain subjective judgements of the similarity of percepts,
- obtain population-state representations,
- compare the two distance matrices.
We test:
D_phenomenal
~
D_neural.
13.46 Experiment Q2 – bistable perception
Use a constant ambivalent stimulus.
Measure:
perceptual report Q_A / Q_B
and at the same time:
neural state M_A / M_B.
Strong evidence will arise if the neural state transition systematically precedes or accompanies the phenomenal transition without a change of the stimulus.
13.47 Experiment Q3 – phase perturbation
Experimentally change the temporal/phase organization with a minimal change of overall activity.
Track:
percept identity,
percept stability,
report.
If the quale-like report changes systematically with the phase geometry, this supports its functional relevance.
13.48 Experiment Q4 – recurrence perturbation
Temporarily disrupt the recurrent integration.
Track the difference between:
stimulus detection
and:
conscious percept report.
If detection persists but the integrated percept does not, recurrence may be causally important for the conscious state.
13.49 Experiment Q5 – persistence threshold
Vary the duration of the stimulus:
Δt.
Look for the threshold at which:
local processing exists
but:
stable reported percept does not.
Compare with the time needed for the formation of a metastable macrostate.
13.50 Experiment Q6 – workspace dissociation
Separate:
local perceptual dynamics
from:
global access/report.
This experiment will not by itself decide the question of phenomenality.
It will, however, help to test the models A/B/C described above.
13.51 Experiment Q7 – hysteresis correlation
Measure simultaneously:
neural hysteresis
and
perceptual hysteresis.
If their dynamic parameters agree systematically, this supports a relation between the manifold state and the conscious percept.
13.52 Experiment Q8 – context-dependent quale geometry
Present the same physical stimulus in a different context.
Measure:
reported phenomenal change
and:
manifold deformation.
Test whether the change of the subjective percept corresponds better to the internal macrostate than to the input itself.
13.53 Cognia experiment CQ1 – functional qualia surrogate
Because Cognia cannot independently report experience, we define only a functional surrogate:
Deep Percept State.
It must satisfy:
persistence,
integration,
context dependence,
causal influence,
prediction,
global accessibility.
It must not be designated as proof of qualia.
13.54 Cognia experiment CQ2 – same input, different internal experience
Cognia receives:
identical X
after two different histories:
H_A
H_B.
If there arise:
M_A != M_B
and the system subsequently:
predicts differently,
acts differently,
reports internal state differently,
we have a functional analogy of a different internal percept.
13.55 Cognia experiment CQ3 – same output, different internal state
A very important experiment:
Two situations lead to the same behavioral output:
Y.
But:
M_A != M_B.
This shows that the internal state contains more information than the behaviour itself.
It is important for the argument that a percept cannot be reduced to an output.
13.56 Cognia experiment CQ4 – same macrostate, different microstate
Create different stochastic realizations:
S1,
S2,
S3.
If all of them belong to:
M_A
and produce the same integrated functional properties, this supports the macroscopic invariance of the Deep Percept.
13.57 Cognia experiment CQ5 – artificial perturbation of the manifold
Artificially move the system:
M_A -> M_B
without a change of the sensory input.
If the following change at the same time:
prediction,
action,
report,
valuation,
then the internal dynamic state genuinely functions as a common perceptual substrate.
13.58 What Cognia cannot demonstrate
Even if all the preceding experiments succeed, it will not be legitimate to claim:
Cognia is conscious.
We may claim only:
Cognia implements and experimentally demonstrates the mechanistic properties that DPSH assumes as a candidate substrate of the phenomenal percept.
13.59 What would increase the plausibility of the phenomenal hypothesis
The plausibility of level E3 would grow if the same dynamic model explains, in biological systems:
subjective reports,
perceptual transitions,
perturbation effects
and at the same time generates in Cognia a corresponding functional architecture.
There would then exist:
cross-substrate mechanistic correspondence.
Still, however, not a direct proof of the subjectivity of an artificial system.
13.60 Falsification criteria
The strong phenomenal hypothesis of DPSH will be weakened if:
- conscious percepts can be stably preserved with the complete removal of the assumed metastable dynamic structures,
- phenomenal changes do not systematically correlate with changes of the relevant manifold,
- phase/timing perturbations strongly change the candidate macrostate without any corresponding change of the percept,
- the same phenomenal state requires a neuronal mechanism fundamentally incompatible with DPSH,
- global or recurrent dynamic mechanisms can be removed without an effect on phenomenal content,
- an alternative simpler model explains the biological data better.
Even a positive result of these tests, however, does not demonstrate an ontological identity of dynamics and qualia.
13.61 The phenomenal hypothesis as the highest layer of DPSH
DPSH therefore has three epistemic levels E1–E3. These are not the
partial hypotheses H1–H13 of the individual chapters, but levels of
claim at which the theory can be assessed:
E1 – Mechanistic
Can the proposed neuronal mechanisms create metastable internal dynamic states?
This can be directly simulated and falsified.
E2 – Perceptual
Can these states represent sensory content, persist, predict and causally govern behaviour?
This can also be tested experimentally.
E3 – Phenomenal
Are these dynamic states at the same time the substrate of subjective experience?
This can be tested only indirectly and remains the strongest open part of the hypothesis.
13.62 Chapter research hypothesis
We formulate the partial hypothesis H12:
H12 – Phenomenal Dynamic Substrate Hypothesis
If the phenomenal properties of a conscious percept supervene on neuronal dynamics, the candidate substrate of these properties may be a temporally extended, integrated and metastable macrostate of a distributed neuronal system rather than an individual neuron, a static activation or a symbolic representation. Phenomenal identity may correspond to a macroscopic dynamic organization that is invariant with respect to part of the microscopic spike variability.
A stronger version:
The geometry and the transition structure of phenomenal experience may correspond systematically to the geometry and the dynamics of the Perceptual Manifold.
This hypothesis is not, however, equivalent to the claim:
metastability = qualia
nor:
Cognia with DPSH = conscious system.
The correct formulation remains:
DPSH mechanisms
->
Deep Percept
->
candidate phenomenal substrate
?
->
subjective experience.
The question mark between the functional dynamic state and subjective experience is part of the theory and must not be removed merely on the basis of the success of a computational model.
13.63 The main epistemic boundary
The strongest scientific claim that a first implementation of DPSH can legitimately make is:
A neuronal architecture has been created in which distributed autonomous units without a global update clock create, through stochasticity, temporal organization, recurrence, plasticity and predictive constraints, persistent, integrated, history-dependent and causally effective internal perceptual states.
If experiments confirm this result, it will constitute support for the Deep Percept mechanism.
The question:
Is the existence of such a state at the same time the existence of phenomenal experience?
will remain an open problem whose resolution requires further empirical as well as theoretical arguments.
14. Integrated predictions and the falsification program for the entire DPSH
14.1 The purpose of the integration chapter
The preceding chapters formulated the individual mechanisms of the Dynamic Perceptual State Hypothesis separately.
This was necessary so that each mechanism could be tested experimentally on its own.
The whole of DPSH does not, however, claim that a conscious or deep percept arises only from:
stochasticity
or:
oscillations
or:
metastability.
The central claim is integrative.
DPSH assumes that the relevant perceptual dynamics arises from the interaction of several mechanisms:
globally unclocked dynamics
+
spontaneous stochastic activity
+
endogenous temporal organization
+
non-commutative event ordering
+
recurrent self-organization
+
symmetry breaking
+
metastability
+
hysteresis
+
predictive constraints
+
local plasticity
+
shared perceptual manifold
+
global accessibility.
It is therefore necessary to test not only:
does mechanism X work?
but also:
does mechanism X causally contribute to the integrated system?
14.2 The hierarchy of hypotheses
The individual partial hypotheses of DPSH do not have the same epistemic strength.
They can be divided into four layers. The names given are binding and
correspond to the declarations in the Chapter research hypothesis
sections of the individual chapters.
Layer A – elementary dynamic mechanisms
This includes:
H1 Globally Clockless Dynamics Hypothesis
H2 Functional Stochasticity Hypothesis
H3 Endogenous Temporal Organization Hypothesis
H4 Non-Commutative Neural Dynamics Hypothesis.
These hypotheses describe the properties of the elementary dynamics of the system.
Layer B – the emergence of the internal macrostate
This includes:
H5 Self-Organized Symmetry Breaking Hypothesis
H6 Metastable Perceptual Manifold Hypothesis
H7 Perceptual Continuity and Hysteresis Hypothesis
H8 Predictively Constrained Dynamics Hypothesis.
These hypotheses describe the formation and maintenance of a functional internal state.
Layer C – learning and system integration
This includes:
H9 Deep State Learning Hypothesis
H10 Global Accessibility Hypothesis
H11 Shared Perceptual Manifold Hypothesis.
These hypotheses describe how the dynamics is learned, shared and made accessible.
Layer D – phenomenal interpretation
This includes:
H12 Phenomenal Dynamic Substrate Hypothesis.
This layer is not directly verifiable by means of Cognia alone.
It depends on the results of the preceding layers and on biological data.
Outside the hierarchy – the technical condition of testability
Chapter 15 formulates one further partial hypothesis:
H13 Experimental Realizability Hypothesis.
H13 is not a superstructure above layer D. It is a precondition without which layers A–D cannot be experimentally tested at all.
These epistemic layers A–D must be distinguished from the three
levels of claim E1–E3 introduced in section 1.12, which describe how
strong a claim about a result can be made at all.
14.3 Dependencies among the hypotheses
Some hypotheses can be tested independently.
Others make sense only if the preceding layers have been at least partially supported.
The working dependency:
H1
|
v
H2 ----+
| |
v v
H3 --> H4
\ /
\ /
v
H5
|
v
H6
/ \
v v
H7 H8
\ /
\ /
v
H9
/ \
v v
H10 H11
\ /
\ /
v
H12.
This diagram is not a claim that all the mechanisms are necessary.
It is an experimental roadmap.
14.4 The minimal core of DPSH
The smallest set of mechanisms that it makes sense to test as the basis of a dynamic percept is:
autonomous stateful units
+
recurrence
+
spontaneous stochasticity
+
temporal organization
+
history dependence.
From this core it must be possible to create:
persistent internal macrostate.
If this is not possible, the higher layers of the theory have no sufficient basis.
14.5 The first integrative prediction
DPSH predicts that a system containing all the basic mechanisms will exhibit a qualitatively different state-space dynamics than the corresponding simplified control networks.
The comparison:
full DPSH core
versus:
synchronous rate-based control.
To measure:
state dimensionality,
metastability,
transition entropy,
history dependence,
perturbation response,
temporal coding,
contextual persistence.
The strong prediction:
dynamics_full
!=
dynamics_control
even with a similar:
task accuracy.
14.6 Accuracy is not the main criterion
An important methodological rule of the whole of DPSH:
A higher classification accuracy does not by itself confirm the hypothesis.
DPSH is a theory of internal dynamics.
A system may have the:
same task accuracy
but a fundamentally different:
internal organization.
The following must therefore always be measured as well:
trajectory structure,
state persistence,
temporal dependency,
causal state influence.
14.7 The ablation program
It must be possible to remove every mechanism separately.
The basic full model:
async
+ stochastic
+ oscillatory
+ recurrent
+ STDP
+ predictive
+ continuous state.
We then create:
- stochasticity
- oscillations
- recurrence
- STDP
- prediction
- history
- workspace
and measure the change of the system.
This provides a causal map of the mechanisms.
14.8 Factorial design
Simple ON/OFF experiments need not be sufficient.
Some mechanisms may function only in mutual interaction.
For example:
stochasticity x oscillations
oscillations x STDP
phase x delays
recurrence x inhibition
prediction x hysteresis.
It is therefore appropriate to use a factorial design.
For example:
stochasticity:
ON / OFF
phase structure:
intact / scrambled
plasticity:
ON / OFF.
This gives:
2 x 2 x 2
experimental conditions.
The interaction effect can thus be measured.
14.9 Synergy of mechanisms
DPSH assumes the possibility that:
effect(A + B)
>
effect(A)
+
effect(B).
For example:
stochasticity alone
->
variability.
oscillation alone
->
temporal regularity.
But:
stochasticity + oscillation
->
structured exploration.
Such a synergic effect is more important for the integrative hypothesis than the individual efficacy of the components.
14.10 Redundancy of mechanisms
Conversely, an experiment may show:
A is not necessary
because:
B can substitute A.
Explicit oscillator cells, for example, need not be necessary if a recurrent network creates an emergent oscillation.
In such a case the hypothesis must be generalized.
Not:
oscillator cell is required
but:
temporal organization is required.
14.11 Necessity versus sufficiency
Every mechanism must be assessed in two directions.
Necessity
If we remove:
X,
does the relevant property disappear?
Sufficiency
If we have only:
X,
does the relevant property arise?
For example:
oscillations
may be:
useful but not sufficient.
DPSH must not confuse these two questions.
14.12 The causal hierarchy
For an individual mechanism one may define:
mechanism
->
intermediate effect
->
macrostate change
->
behavioral consequence.
For example:
phase relation
->
transmission efficacy
->
state transition
->
perceptual choice.
An experiment must ideally measure all the intermediate steps.
Otherwise there is a risk of an incorrect attribution of causality.
14.13 The primary integrative hypothesis
The central mechanistic hypothesis of the whole of DPSH can be formulated as:
In a globally clockless recurrent network, the interaction of spontaneous stochasticity, local temporal organization, history-dependent nonlinear dynamics, recurrent selection and local plasticity may create metastable distributed macrostates whose geometry preserves the perceptual context, generates predictions and causally influences further processing.
This is the main testable thesis.
14.14 The strong integrative prediction
If DPSH is right, it must hold that:
remove temporal organization
->
degrade macrostate structure
even if we approximately preserve the:
firing rate,
spike count,
network size,
sensory information.
Likewise:
remove recurrence
->
degrade persistence.
And:
remove history
->
degrade context dependence.
That is, different mechanisms must have specific failure signatures.
14.15 Failure signatures
Every mechanism should have an expected type of failure.
Without stochasticity
We expect possible:
rigidity,
poor exploration,
reduced spontaneous transitions.
Without phase organization
We expect:
poorer temporal coordination,
weaker dynamic routing,
degraded timing-sensitive learning.
Without recurrence
We expect:
poor persistence,
reduced metastability.
Without history
We expect:
loss of hysteresis,
reduced context dependence.
Without prediction
We expect:
excessive drift
or
poor adaptation to changed environment.
Without plasticity
We expect:
fixed manifold geometry.
Such signatures are important for falsification.
14.16 Global prediction no. 1 – microscopic variability, macroscopic stability
DPSH predicts:
spike patterns vary
while:
macrostate identity persists.
That is:
high microstate variability
+
low macrostate variability.
If a stable percept requires almost identical spike patterns, this part of the hypothesis will be weakened.
14.17 Global prediction no. 2 – timing matters beyond rate
The strong prediction:
same approximate rate
same spike count
same topology
but:
different temporal organization
leads to:
different internal dynamics.
If not, the temporal part of DPSH loses its significance.
14.18 Global prediction no. 3 – history changes the present
For:
same current input X
after:
history A
and:
history B
there must exist:
S_A(X) != S_B(X)
at least in some relevant tasks.
This is the basic condition of a continuous internal model.
14.19 Global prediction no. 4 – perception survives brief input loss
After the formation of:
M_A
a brief:
sensory dropout
must not immediately destroy the whole perceptual state.
There must be a measurable persistence.
14.20 Global prediction no. 5 – internal state affects ambiguous input
After:
M_A
and:
M_B
the same:
X_ambiguous
must lead to different probabilities of the outcome.
This will demonstrate the functional significance of the internal state.
14.21 Global prediction no. 6 – learned environment changes spontaneous dynamics
After learning:
spontaneous_before
!=
spontaneous_after.
The stronger prediction:
spontaneous_after
will be structurally more similar to the learned evoked states.
14.22 Global prediction no. 7 – prediction stabilizes reality-consistent states
A state with a low prediction error:
M_correct
is to have a higher persistence than:
M_inconsistent
under otherwise comparable conditions.
14.23 Global prediction no. 8 – excessive prediction produces pathological persistence
With too strong a:
top-down gain
we expect:
reduced sensory correction,
excessive state persistence.
This is a systemic prediction of the balance between the internal model and reality.
14.24 Global prediction no. 9 – percept formation and global access are separable
There must exist at least some conditions:
local percept present
global access reduced.
If the two processes cannot be separated, H10 will have to be reformulated.
14.25 Global prediction no. 10 – shared state supports multiple functions
The same internal macrostate should be usable for:
prediction,
action,
valuation,
report.
If every module needs a completely independent representation, H11 is weakened.
14.26 Experimental phase 0 – validation of the engine
Before we begin to test DPSH, it must be verified that Cognia correctly implements the basic physics of the model.
To test:
event ordering,
delays,
stochastic distribution,
oscillator phase,
refractory periods,
plasticity timing,
state logging.
If the engine is not deterministically reproducible with a fixed random seed, it will not be possible to interpret the results.
14.27 Reproducibility
Every experiment must support:
random seed,
network snapshot,
exact configuration,
event log.
The experiment:
run(configuration, seed)
must be repeatable.
14.28 Frozen randomness control
For stochastic experiments it is necessary to compare:
same network
same input
same stochastic sequence
against:
same network
fresh stochastic sequence.
In this way it is possible to separate the:
stochastic distribution effect
from the:
exploration effect.
14.29 Experimental phase 1 – autonomous dynamics
The first goal is not perception.
It is:
Can the network create non-trivial dynamics without external input?
To test:
deterministic silence,
spontaneous activity,
oscillatory regimes,
metastability.
If not, the further layers do not yet make sense.
14.30 Experimental phase 2 – emergence of metastability
The network receives simple competitive conditions:
A
B.
We test:
symmetry breaking,
state persistence,
spontaneous switching.
We measure:
order parameter,
dwell time,
transition entropy.
14.31 Experimental phase 3 – temporal causality
We test:
A -> B
versus:
B -> A.
And:
phase intact
versus:
phase scrambled.
This will verify whether timing genuinely constitutes an important state variable.
14.32 Experimental phase 4 – perceptual context
We use:
A -> blank -> X_ambiguous
and:
B -> blank -> X_ambiguous.
This is the first genuine test of the Deep Percept.
The condition of success is:
internal state during blank
->
predicts later choice.
14.33 Experimental phase 5 – causal perturbation
It is not sufficient to decode the state.
We must actively change it.
For example:
M_A -> perturb -> M_B.
Then:
same X
must lead to a different result.
This is fundamental evidence of causality.
14.34 Experimental phase 6 – predictive world model
The network receives a temporally structured environment.
We test:
next-state prediction,
occlusion,
unexpected continuation,
correction after mismatch.
Here the memory state becomes a model of the environment.
14.35 Experimental phase 7 – Deep State Learning
Only after a stable manifold has been created do we switch on:
ongoing plasticity during spontaneous activity.
We test:
consolidation,
generalization,
drift,
self-reinforcement.
This is the risky part of the hypothesis and must be tested with caution.
14.36 Experimental phase 8 – the multimodal manifold
We add more input modalities.
For example:
visual-like input
audio-like input.
We test:
cross-modal completion,
shared state,
context integration.
14.37 Experimental phase 9 – Global Workspace
Only after that do we add:
workspace.
We test:
global broadcast,
competition,
ignition,
top-down feedback.
The workspace must not mask a failure of the lower perceptual layer.
14.38 Experimental phase 10 – the phenomenal analogy
Cognia cannot directly test:
qualia.
It can, however, test:
Deep Percept properties.
The biological literature must subsequently verify whether the same dynamic mechanisms correlate and are causally related to the reported conscious percept.
14.39 Stop conditions
An important part of the program are the conditions under which the research is not simply to move on.
For example:
Stop A
If phase scrambling has no effect in a rate-matched control, do not use phase as a central mechanism in further experiments.
Stop B
If a persistent state does not exist without an explicit memory cell, revise H6.
Stop C
If a history reset does not change behaviour, revise H7.
Stop D
If predictive feedback does not change adaptation, revise H8.
Such a procedure prevents the theory from becoming unfalsifiable.
14.40 Criteria of success for the first generation of Cognia DPSH
We would not regard the following as the first significant success:
"system behaves intelligently."
A minimal experimental package should show simultaneously:
- autonomous ongoing activity,
- metastable population states,
- persistence across a sensory blank,
- history-dependent interpretation,
- phase/timing causal effect beyond rate,
- causal effect of state perturbation,
- predictive continuation,
- learning-induced state-space deformation.
That would constitute strong support for the Deep Percept mechanism.
14.41 Criteria of a stronger success
A stronger system should in addition show:
- spontaneous replay,
- useful ongoing plasticity,
- multimodal integration,
- shared downstream projections,
- dynamic routing,
- global access competition,
- workspace feedback.
Such a system would be a very interesting implementation of the whole functional layer of DPSH.
14.42 Negative results are part of the theory
If a mechanism fails, this is not a failure of the research.
For example:
oscillator cells not needed
may lead to a better theory:
emergent temporal structure is sufficient.
Or:
stochasticity not necessary
may mean that the relevant exploration arises by another mechanism.
DPSH must be willing to remove its own components.
14.43 Model reduction
After every experimental cycle it is appropriate to look for the simplest model that still reproduces the relevant phenomena.
That is:
full model
->
remove unnecessary mechanism
->
simpler explanatory core.
The aim is not to maximize the number of interesting mechanisms.
The aim is to find the minimal causal architecture.
14.44 Competing models
For every important experiment there must exist at least one simpler alternative model.
For example:
DPSH model
dynamic metastable state.
Alternative A
explicit memory register.
Alternative B
conventional RNN hidden state.
Alternative C
rate-coded attractor.
If all of them explain the data equally well, DPSH has no sufficient explanatory advantage.
14.45 Model comparison
To compare not only:
task accuracy.
Also:
complexity,
robustness,
generalization,
state richness,
perturbation behavior,
temporal sensitivity.
A stronger model must explain more of the relevant phenomena.
14.46 Falsification of the strong version of DPSH
The strong mechanistic version of DPSH will be seriously weakened if it turns out that:
- a synchronous control creates the same relevant dynamic properties,
- stochasticity brings no specific effect,
- timing and phase can be removed without a loss of function,
- the order of events is irrelevant,
- metastability is not needed,
- history dependence brings no effect,
- simple explicit memory explains all the perceptual results,
- prediction is not needed for reality grounding,
- a shared manifold cannot be functionally demonstrated,
- the workspace is not separable from the basic percept.
In such a case it would be necessary to reduce the original theory fundamentally.
14.47 Falsification of the phenomenal version
The phenomenal H12 will be weakened if biological experiments show that:
conscious percept
systematically persists without the dynamic mechanisms that DPSH regards as a candidate substrate.
Dissociations would be especially important:
same phenomenal state
+
radically different relevant macrostate
or:
same macrostate
+
systematically different phenomenal state.
Such results would weaken the structural correspondence.
14.48 What DPSH would not regard as a confirmation
The following alone cannot be regarded as a confirmation:
high classification accuracy,
human-like text output,
self-report of consciousness,
complex behavior,
large neural network,
presence of oscillations,
presence of metastability.
Each of these properties may exist without the central mechanism of DPSH.
14.49 The most important confirming result
We would regard the following situation as an exceptionally strong result:
same network
same sensory input
same approximate firing rate
same spike count
same connectivity
but:
different relative timing / phase organization
leads to:
different metastable perceptual state
and this change:
causally changes later interpretation.
This would provide very direct support for the claim that temporal organization is not a secondary detail, but part of the internal representation.
14.50 The second strongest result
A further strong result:
sensory input removed
but:
internal context persists,
and:
perturbing that internal state
changes the response to a later:
identical ambiguous input.
We would thereby demonstrate the functional existence of an internal dynamic state beyond the current sensory stream.
14.51 The third strongest result
The third:
learning
->
spontaneous dynamics changes
and spontaneous learning:
improves held-out generalization.
This would provide support for the strong Deep State Learning hypothesis.
14.52 Integrated Deep Percept criterion
For the purposes of the research we define a summary criterion.
A system has a Deep Percept if the internal state simultaneously satisfies:
D1 decodability
D2 persistence
D3 metastability
D4 history dependence
D5 temporal sensitivity
D6 predictive relevance
D7 causal downstream influence
D8 robustness
D9 generalization
D10 distributed integration.
It is not necessary that every criterion be binary.
A vector can be created:
DP =
(
d1,
d2,
...,
d10
).
This will make it possible to compare different architectures without a single arbitrary "consciousness score".
14.53 Deep Percept Index
For practical experiments a composite metric may later be created:
DPI = F(d1, d2, ..., d10).
It is important, however, that:
DPI
not be interpreted as:
degree of consciousness.
It is merely a technical metric of the functional properties of the Deep Percept.
14.54 The research register
Every experiment should be recorded in the format:
hypothesis
prediction
architecture
control
manipulated variable
dependent metrics
result
falsification status
interpretation.
This makes it possible to prevent the retrospective adaptation of the hypothesis to the results.
14.55 The pre-registration principle
For key experiments it is appropriate to write down explicitly, before the run:
expected outcome,
null outcome,
falsifying outcome.
For example:
phase scramble
->
predicted decrease in state separability.
If the result does not occur, it must be recorded as negative.
14.56 Experimental versions of DPSH
The theory should have versions:
DPSH 0.1
DPSH 0.2
...
Every version records:
retained hypotheses,
rejected hypotheses,
modified mechanisms.
The Cognia implementation must be versioned together with the theory.
14.57 The relation between the theory and the Cognia engine
The development of the engine must not run ahead of the hypothesis in such a way that a mechanism is added only because it improves performance.
Every significant property of the engine must have the link:
theoretical assumption
->
implementation
->
experimental prediction.
For example:
H3 temporal organization
->
local oscillator API
->
phase scramble experiment.
14.58 Cognia as experimental apparatus
Cognia here is not merely the resulting AI.
It is above all an:
experimental platform.
It must allow:
mechanism isolation,
precise perturbation,
logging,
replay,
ablation,
state-space analysis.
This is more important than an immediate ability to solve complex tasks.
14.59 The priority of small networks
The first experiments should use the smallest networks in which the phenomenon is observable.
The advantages:
interpretable dynamics,
cheaper parameter sweeps,
easier causal analysis,
fewer confounds.
A large network makes sense only once the mechanism works in a small one.
14.60 Scaling experiments only later
Only after the basic mechanisms have been demonstrated is it appropriate to vary:
N neurons,
connectivity density,
oscillator count,
delay distribution.
It is then possible to test:
scaling laws.
For example:
state richness vs N
robustness vs redundancy
connectivity vs timing structure.
14.61 The main research roadmap
The whole program can be simplified:
Stage 1
Can autonomous dynamics exist?
Stage 2
Can it self-organize metastable states?
Stage 3
Does timing causally matter?
Stage 4
Can a state retain perceptual context?
Stage 5
Does that state causally alter future interpretation?
Stage 6
Can it predict the environment?
Stage 7
Can experience reshape the state space?
Stage 8
Can spontaneous dynamics consolidate it?
Stage 9
Can multiple functions share the state?
Stage 10
Can selected content become globally available?
Stage 11
Do corresponding biological dynamics track phenomenal perception?
14.62 The minimum publishable experiment
The first publication need not verify the whole of DPSH.
On the contrary, it should test one central causal thesis.
The strongest candidate:
The relative temporal organization of stochastic spiking activity causally contributes to the formation and maintenance of a metastable perceptual state beyond the firing rate itself.
The experiment:
A -> blank -> ambiguous X
in a:
phase-intact
and a:
rate-matched phase-scrambled
network.
To measure:
state separability,
persistence,
causal behavioral effect.
14.63 The second publication
If the first hypothesis succeeds:
Spontaneous ongoing activity and local plasticity change the geometry of the learned state space and may improve future perceptual prediction.
Here Deep State Learning would be tested:
Deep State Learning.
14.64 The third publication
A further one:
A metastable perceptual state may function as a shared internal model for several downstream subsystems before entry into the Global Workspace.
The Perceptual Manifold would thereby be tested.
14.65 The phenomenal publication only later
Only after the mechanistic part has been supported does it make sense to publish the broader theoretical argument:
Deep Percept
->
candidate phenomenal substrate.
Without a functional mechanism the phenomenal part would be too speculative.
14.66 The central falsification question of the whole of DPSH
The whole mechanistic theory can be summarized by the question:
Does the relative temporally organized trajectory of a distributed autonomous neuronal network contain causal information about its internal perceptual state that cannot be reduced to the current sensory input, to firing rates or to an explicit memory variable?
If:
no,
then a large part of DPSH loses its reason for existing.
If:
yes,
the following question ensues:
Can this dynamic structure be continuously learned, predictively anchored in the world and shared among functional subsystems?
14.67 The final structure of the claims
DPSH therefore proceeds from the weakest to the strongest claim:
Claim A
dynamic states exist.
Claim B
dynamic states carry information.
Claim C
dynamic states causally affect future processing.
Claim D
dynamic states form integrated perceptual context.
Claim E
experience reshapes their geometry.
Claim F
they function as shared internal world model.
Claim G
selected content becomes globally available.
Claim H
these mechanisms may constitute a candidate substrate of
phenomenal experience.
Every step must be supported separately.
14.68 The main methodological principle
DPSH must be designed so that it can fail.
If we explain every negative experiment by adding a new auxiliary hypothesis, the theory will cease to be scientifically useful.
Therefore it holds:
A mechanism that repeatedly fails to bring the predicted causal effect must be removed from the central theory or degraded to an optional implementation property.
14.69 The integrated hypothesis of the whole of DPSH
The final working formulation of the mechanistic part reads:
The Dynamic Perceptual State Hypothesis assumes that a continuous internal percept may arise as a metastable macroscopic trajectory of a globally clockless recurrent network of autonomous stochastic units. Relative timing, endogenous oscillatory organization, synaptic delays, non-commutative history, recurrent selection and local plasticity jointly shape a state space whose dynamics preserves the sensory context, generates predictions and influences future interpretation. Sensory evidence continuously constrains this internal model, and selected states may subsequently gain global availability through a workspace mechanism.
The phenomenal extension remains:
If subjective phenomenal content supervenes on neuronal dynamics, such an integrated temporally extended dynamic state constitutes a candidate substrate of phenomenal experience.
The first claim must be experimentally falsified or supported.
The second remains an open theoretical hypothesis.
14.70 The immediate next step
After the formulation of the whole of DPSH, the next step is no longer the addition of further theoretical mechanisms.
It is:
freeze theory version 0.1
and the conversion of the individual hypotheses into:
Cognia engine requirements
->
minimal experimental architectures
->
predefined metrics
->
ablation experiments.
From this moment on, new architectural elements must arise primarily as a response to:
an experimental problem
or:
a falsified part of the hypothesis,
not merely because they intuitively resemble a biological brain.
15. DPSH requirements for the Cognia engine and the experimental framework
15.1 The purpose of this chapter
The Dynamic Perceptual State Hypothesis is an experimental hypothesis.
So that its individual parts can genuinely be tested, the Cognia engine must not function merely as an environment for defining neural networks.
It must at the same time function as a:
simulator,
perturbation framework,
recorder,
replay system,
ablation platform,
experimental runtime.
The basic requirement is:
It must be possible to implement, parameterize, measure and switch off every mechanism assumed by DPSH separately, and to compare it with a control variant.
Cognia is therefore not merely to create a system that "works".
It must make it possible to find out:
why it works,
which mechanism is necessary,
which mechanism is redundant,
what changes after its removal.
15.2 The separation of the three layers of the system
Cognia should consistently separate at least three layers:
1. Engine layer
It provides:
event processing,
simulation time,
random generator,
logging,
snapshotting,
replay.
2. Neural architecture layer
It defines:
neurons,
synapses,
oscillators,
memory units,
modulators,
controllers,
populations.
3. Experimental layer
It defines:
stimuli,
interventions,
ablations,
measurements,
hypotheses,
expected outcomes.
This separation is fundamental.
An experimental condition must not be covertly implemented as a change of the engine itself.
15.3 The global time of the engine versus a global clock of the network
The engine of course needs to represent time in some way.
This does not, however, mean that the neuronal network uses a global clock.
It is necessary to distinguish:
simulation time
from:
neural processing clock.
The engine may record:
t = 12.351 ms
and process an event.
The neuronal architecture must not, however, be forced into:
tick 1 -> update all neurons
tick 2 -> update all neurons
tick 3 -> update all neurons.
That is:
engine has time
but:
network has no mandatory global update step.
15.4 Event-driven execution
The basic regime of Cognia DPSH should be event-driven.
An event may be, for example:
spike,
oscillator phase event,
sensory event,
plasticity event,
refractory end,
modulation event.
Every event has:
source,
target,
timestamp,
type,
payload.
For example:
Event {
source: neuron_12
target: neuron_54
type: spike
time: 12.351 ms
}.
The engine processes events in temporal order.
15.5 Event queue
An explicit:
event queue.
is needed.
Minimal properties:
insert(event),
pop_next(),
peek_next_time(),
cancel(event),
inspect_queue().
The queue must respect:
exact timestamp order.
With an identical timestamp the behaviour must be:
deterministic under fixed seed
or the ordering rule must be explicitly defined.
15.6 Stable ordering at the same time
If two events occur at the same time:
t_A = t_B,
it must be clear whether:
A then B
or:
B then A
or:
both treated as simultaneous.
This is fundamental for non-commutative dynamics.
The engine must not arbitrarily change the order without a record.
15.7 The local lifecycle of a neuron
Every neuron must have its own state:
state_i(t).
At a minimum it may contain:
membrane-like activation,
refractory state,
adaptation,
baseline spike probability,
recent spike history,
local modulation.
A neuron is updated:
when relevant event arrives
or:
when its own internal dynamics requires it.
It need not be recomputed at every change of the global time.
15.8 Autonomous internal events
A neuron must be able to schedule its own future event.
For example:
next spontaneous spike candidate,
refractory end,
adaptation decay.
This enables an autonomous lifecycle without a global tick.
15.9 Stochastic neuron
Cognia must contain at least one parameterizable stochastic neuron model.
It must support:
baseline firing,
input-dependent firing probability,
refractory period,
optional noise amplitude,
random seed control.
In general:
P(spike_i,t) =
F(
baseline,
current_state,
input,
oscillator modulation,
stochastic component
).
15.10 Deterministic control neuron
For every stochastic model there must exist a corresponding deterministic control.
For example:
same input transform,
same threshold,
no random component.
This will make it possible to test directly:
stochastic ON
versus:
stochastic OFF.
15.11 Random subsystem
Randomness must not be scattered uncontrollably throughout the engine.
Cognia must have a centralized experimental random subsystem:
RandomContext.
It must support:
seed,
stream id,
snapshot,
replay.
For example:
random("neuron_12", "spike_generation").
This will make exact reproduction possible.
15.12 Separate random streams
It is appropriate to separate the random streams by function:
spontaneous spikes,
synaptic noise,
stimulus noise,
initialization,
learning noise.
It is then possible, for example, to freeze:
spontaneous randomness
and vary only:
input noise.
15.13 Frozen randomness
The engine must allow:
record random sequence
and subsequently:
replay exact sequence.
This is necessary for experiments comparing:
fresh stochasticity
versus:
frozen stochasticity.
15.14 Synapse
A synapse must not be merely a:
weight.
The minimal state:
Synapse {
weight
delay
plasticity_rule
plasticity_state
enabled
}.
15.15 Synaptic delay
Every synapse must support:
delay_ij.
A spike arising at:
t
arrives at:
t + delay_ij.
The delay must be experimentally variable and loggable.
15.16 Fixed versus adaptive delay
The engine should support two variants:
Fixed delay
d_ij = const.
Plastic delay
d_ij(t)
variable through learning.
The second variant is not necessary for the first experiments, but the architecture should not make it impossible.
15.17 Excitatory and inhibitory synapses
The following must be explicitly supported:
excitatory connection
and:
inhibitory connection.
Inhibition must not be implemented merely as a negative workaround if this makes different plasticity or dynamics impossible.
15.18 Plasticity as a separate mechanism
The plasticity rule must be separated from the neuron and the synapse in such a way that the following can easily be interchanged:
STDP,
Hebbian,
rate-based,
no plasticity.
For example:
plasticity STDP {
pre_window
post_window
learning_rate
}.
15.19 STDP
A minimal STDP implementation must have:
t_pre,
t_post,
Δt,
Δw.
And log:
every plasticity event.
This will make it possible to verify retrospectively whether learning genuinely proceeded according to the expected timing.
15.20 Modulated plasticity
Plasticity must allow modulation:
learning_gain.
For example:
prediction error,
relevance,
reward,
consolidation gate.
Formally:
Δw =
learning_gain * STDP(Δt).
15.21 Learning gate
There must be the possibility of:
plasticity ON
plasticity OFF
both globally and locally.
For example:
group.visual.plasticity = off.
This is fundamental for ablation experiments.
15.22 The oscillator as a network element
An explicit oscillator must be defined as an object of the neuronal architecture.
For example:
oscillator Theta {
frequency: 8 Hz
phase: 0
amplitude: 1
}.
An oscillator is not a scheduler of the engine.
15.23 Oscillator output
An oscillator may generate:
continuous modulation
or:
discrete periodic events.
It must be possible to define its coupling to neurons:
oscillator -> neuron group.
And the type of modulation:
excitability,
threshold,
spike probability,
synaptic gain,
plasticity gain.
15.24 Multiple oscillators
The network must support many independent local oscillators:
O1,
O2,
...
On.
Each with:
frequency,
phase,
amplitude,
coupling.
15.25 Phase query
For every event it must be possible to determine:
phase(O_k, t).
This is necessary for the subsequent analysis of:
spike phase distribution,
phase locking,
phase-dependent plasticity.
15.26 Phase scrambling
The experimental framework must support the operation:
phase_scramble(group).
This may:
randomize oscillator phase,
destroy cross-population phase relationships,
but ideally preserve:
mean oscillation frequency,
amplitude,
average firing rate.
This is one of the key experimental interventions of the whole of DPSH.
15.27 Phase jitter
Alongside complete scrambling there must exist:
phase_jitter(σ_phase).
This will make it possible to look for a temporal threshold:
how much phase disruption breaks function.
15.28 Emergent oscillation
The engine must not require an explicit oscillator object.
It must be possible to create a recurrent microcircuit out of which an oscillation arises emergently.
The recorder must be able to detect its presence subsequently.
15.29 Controller
Cognia may contain controllers.
It is necessary, however, to distinguish a:
modulatory controller
from a:
hidden central executive.
A controller may set:
gain,
attention,
learning gate,
relevance.
It should not, for example, call directly:
percept = choose_best_state().
If it did so, the self-organization would not be genuinely emergent.
15.30 Memory cells
Explicit memory units may exist.
They must, however, be clearly marked as:
explicit state storage.
This will make it possible to compare:
dynamic memory
versus:
explicit memory cell.
15.31 Memory ablation
The experimental framework must allow:
memory_cells OFF
without changing the rest of the architecture.
This is necessary for the test:
does metastable state retain context without explicit memory?
15.32 Population abstraction
Cognia should support working with populations:
population VisualA[100]
population VisualB[100].
It is needed for:
symmetric competition,
excitation/inhibition,
oscillator modulation,
collective analysis.
15.33 Population recorder
For every population it must be possible to log:
spike count,
firing rate,
population state,
oscillator phase,
coherence,
mean activation.
This will make it easier to track macrostates.
15.34 Global state recorder
One of the most important components:
StateRecorder.
It must be possible to store over time:
neuron states,
spike events,
oscillator states,
synaptic weights,
relevant modulatory states.
The object of research is not only the output.
It is:
S(t).
15.35 Sampling the state space
It need not be practical to store the complete state after every micro-event.
The recorder must support:
fixed sampling interval,
event-triggered snapshot,
selected variable recording.
For example:
sample every 1 ms
or:
snapshot on macrostate transition.
15.36 Spike log
The spike log must contain at least:
neuron id,
population,
timestamp,
local phase,
incoming cause if available.
This will make it possible to reconstruct:
spike trains,
order,
causal chains.
15.37 Causal trace
A very useful function:
trace(event_id).
This should show:
which events contributed to this event.
It need not be full philosophical causality.
A technical provenance graph suffices.
15.38 Synaptic change log
Every change of:
w_ij
must have:
old value,
new value,
timestamp,
pre event,
post event,
modulation value,
learning rule.
This is fundamental for Deep State Learning experiments.
15.39 Network snapshot
The engine must support:
snapshot network_state.
The snapshot must include:
neurons,
synapses,
oscillator phases,
random streams,
pending events,
plasticity state.
Only in this way is it possible to create exactly two experimental branches from the same initial state.
15.40 Restore
There must exist:
restore(snapshot).
We can then carry out:
branch A = phase intact
branch B = phase scrambled
from exactly the same moment.
15.41 Experimental branching
The ideal framework:
snapshot S0
/\
/ \
A B
A:
control condition.
B:
intervention.
We compare:
trajectories.
15.42 Ablation API
It should be possible to switch off every significant mechanism without rewriting the architecture.
For example:
ablate stochasticity
ablate oscillators
ablate recurrence
ablate plasticity
ablate prediction
ablate workspace.
This reduces the risk of implementation confounds.
15.43 Soft ablation
Alongside ON/OFF there should also exist gradual manipulation:
recurrence_gain = 0.0 ... 1.0
stochasticity = 0.0 ... X
phase_jitter = 0 ... X.
Many transitions will probably be nonlinear.
15.44 Matched controls
The framework must make it possible to create controls actively with preserved statistics.
A phase scrambling experiment, for example, must ideally maintain a similar:
firing rate,
spike count,
total activity.
This may require:
adaptive gain normalization.
15.45 Rate-matched control
For experiments with timing there should exist a helper:
match_firing_rate(reference, target).
This makes it possible to reduce the possibility that a difference arose merely because of a change in the amount of activity.
15.46 Spike-count-matched replay
A further possibility:
record spike train
and subsequently create a:
reordered train
with the same:
neuron participation,
spike count.
We manipulate only the timing/order.
15.47 Order scrambling
The framework must support:
reorder_events(window, mode).
For example:
reverse,
random permutation,
fixed jitter.
This is key for the test of non-commutative dynamics.
15.48 Delay perturbation
It must be possible to:
perturb_delays(group, distribution).
For example:
+1 ms jitter
shuffle delays
zero delays.
This will make it possible to test the temporal topology.
15.49 Input framework
A stimulus must not be hardcoded into the architecture.
There must exist:
InputSource.
This generates temporally defined sensory events.
15.50 Stimulus sequence
For example:
stimulus A from 0–100 ms
blank 100–300 ms
ambiguous X 300–400 ms.
The experiment must be written declaratively.
15.51 Ambiguous input
The framework must support inputs that:
equally support multiple states.
This is necessary for:
symmetry breaking,
hysteresis,
perceptual reversal.
15.52 Continuous stimulus sweep
For hysteresis it must be possible to generate a parametric input:
x(t) = 0 -> 1
and then:
1 -> 0.
With an exact log of the transition threshold.
15.53 Occlusion
The sensory framework must support:
temporary input removal
without a reset of the network.
This is important for the test of continuity.
15.54 Unexpected continuation
For predictive experiments:
learned A -> B
but the test:
A -> C.
The framework must record exactly:
expected sequence
actual sequence.
15.55 Multi-modal inputs
Later it must be possible to define parallel:
visual input
audio input
body input.
Each with its own timestamping.
15.56 The output is not the only metric
The experiment framework must not assume that every trial has:
one output label.
The output may be:
action,
trajectory,
state classification,
transition time,
prediction error.
15.57 State-space export
Cognia must be able to export experimental data into a format suitable for:
PCA,
UMAP,
clustering,
transition analysis,
trajectory comparison.
For example:
CSV,
JSONL,
binary matrix.
15.58 Feature selection
We need not analyse every internal parameter.
The framework should allow:
record feature set.
For example:
neuron activation only
or:
activation + phase + synaptic variables.
15.59 Macrostate detector
For some experiments a module will be useful:
MacrostateDetector.
It should not be part of the network.
It is merely an analytical tool of the observer.
It may, for example, identify:
clusters,
transitions,
dwell times.
15.60 No hidden percept object
The engine must not contain a state:
CurrentPercept = A
unless the neuronal system itself has created it as an explicit downstream representation.
An analytical tool may later say:
trajectory classified as M_A.
But that is external analysis.
15.61 Decodability test
The framework must make it possible to train an external decoder over the state data:
state -> context label.
The decoder must not influence the network itself.
It is merely a measuring instrument.
15.62 Causal perturbation
Alongside decoding it must be possible to intervene in the state space.
For example:
stimulate population A
inhibit population B
shift oscillator phase.
And then track:
trajectory change,
behavior change.
15.63 Perturbation API
For example:
perturb {
at: 250 ms
target: population.A
type: activation
strength: 0.2
}.
15.64 Local perturbation
Interventions should be local as far as possible.
A global:
set_state(M_B)
would violate the principle of emergent dynamics.
Better:
activate subset,
inhibit subset,
phase shift,
synaptic perturbation.
15.65 Workspace layer
The Global Workspace must be implemented as a separate dynamic structure.
Not as a:
global variable.
It should have:
workspace populations,
recurrent connections,
candidate inputs,
broadcast outputs.
15.66 Workspace access measurement
The framework must measure:
which modules received information,
when,
for how long.
This will make it possible to define:
global accessibility.
15.67 Workspace ablation
It must be possible to:
disable workspace
without removing the lower perceptual modules.
15.68 Workspace feedback
It must be possible to separate experimentally:
feedforward access
and:
top-down feedback.
For example:
broadcast ON
feedback OFF.
15.69 Predictive subsystem
Cognia must support at least two possibilities.
Explicit predictor
state -> predicted sensory input.
Implicit prediction
learned state transitions.
Both must be separately testable.
15.70 Prediction-error channel
If we use an explicit prediction error, it should be distributable.
Not merely:
global_error scalar.
For example:
error.visual.position
error.visual.shape
error.audio.frequency.
15.71 Precision modulation
A later version must allow:
error gain.
For example:
precision = 0.2
for a noisy sensor.
In this way the attention/precision hypotheses can be tested.
15.72 The controller as a modulator of prediction
A controller may, for example, change:
prediction gain,
sensory gain,
attention.
But it should not directly set:
correct percept.
15.73 Experiment definition
Every experiment should be describable declaratively.
For example:
experiment PhaseScramble {
hypothesis: H3
seed: 42
stimulus: A_blank_X
branches:
control:
phase: intact
intervention:
phase: scrambled
metrics:
state_separability
dwell_time
choice
}.
15.74 Hypothesis metadata
Every experiment should have:
hypothesis id,
prediction,
null hypothesis,
falsification criterion.
This directly connects the theory with the result.
15.75 Predefined outcome
For example:
prediction:
phase scramble decreases state separability
null:
no significant change
falsification:
no effect across predefined parameter range.
This limits post-hoc interpretation.
15.76 Experiment registry
The framework should record:
experiment id,
code version,
network version,
theory version,
seed,
date,
parameters,
result.
This is fundamental for reproducibility.
15.77 Theory version
Every experiment must record:
DPSH version.
For example:
DPSH-0.1.
If we later change the hypothesis, we must not retrospectively reinterpret an old experiment without marking it.
15.78 Engine version
Similarly:
Cognia engine version.
A result must be reproducible against a particular runtime.
15.79 Network configuration hash
Every network may have a:
config hash.
This makes it possible to verify that two conditions genuinely use the same architecture.
15.80 Parameter sweeps
The framework must make it possible to vary automatically:
stochasticity,
recurrence,
oscillator frequency,
phase,
delay,
learning rate.
The output:
parameter -> metric.
15.81 Multidimensional sweep
For example:
stochasticity x phase jitter x recurrence gain.
This will be important for the search for:
critical regimes.
15.82 Do not look only for the best result
A sweep must not be used only to find the:
highest score.
We must analyse:
regime transitions,
stability regions,
failure boundaries.
15.83 Criticality detection
If, for example, at:
g_rec = g_c
there is a sudden:
state coherence increase,
the framework should make it possible to capture:
phase transition-like behavior.
15.84 Replication across seeds
Every stochastic experiment must be run:
across multiple seeds.
The result is not a single run.
It is a distribution:
P(metric).
15.85 Confidence intervals
The experimental output must support:
mean,
variance,
confidence interval.
Without this it will not be possible to separate an effect from stochastic variability.
15.86 Baseline models
The Cognia framework should support control architectures:
synchronous RNN-like model,
explicit memory model,
rate-coded attractor,
deterministic recurrent model.
DPSH must be compared with simpler alternatives.
15.87 Synchronous control
An especially important control:
same architecture
but:
synchronous batched updates.
The aim is to test the difference between:
event-driven timing
versus:
global discretization.
15.88 Rate-based control
A further one:
convert spike activity to rates.
This will make it possible to test whether timing contains information beyond rate.
15.89 Explicit memory control
For example:
memory bit stores context A/B.
If this simple variant explains all the results equally well, DPSH must demonstrate a different advantage.
15.90 Analytical observer separation
All metrics such as:
cluster id,
manifold dimension,
macrostate label
must exist outside the neuronal network.
The network itself must not receive this analytical information unless that is explicitly part of the experiment.
15.91 Performance profiler
Because an event-driven system may be computationally demanding, the engine must measure:
events per second,
queue size,
memory usage,
neuron updates,
synaptic events.
This will make it possible to address scaling later.
15.92 Sparse connectivity
The engine should be optimized from the outset for a:
sparse graph.
This is important both biologically and computationally.
It must not assume:
all-to-all connectivity.
15.93 A large number of simple neurons
The architecture must allow the experiment:
many simple neurons
versus:
fewer complex neurons.
This is important for the later scaling hypothesis.
15.94 Dynamic creation of populations
Later experiments may require:
dynamic recruitment
of a neuron into different assemblies.
The engine must therefore not firmly assume that the functional role of a neuron is unchangeable during a run.
15.95 Neuromodulation
A later version should support global or regional modulators:
dopamine-like,
relevance,
arousal,
learning gain.
Not as a biological copy, but as a general modulatory signal.
15.96 Internal input
Cognia must support the difference between:
external sensory input
and:
internal neural input.
Internal modules must be able to enter the same dynamic system as the sensory channels.
This is important for:
memory,
valuation,
attention,
prediction,
workspace feedback.
15.97 No privileged external/internal API
From the point of view of a neuron it may be appropriate for sensory and internal events to have a similar delivery mechanism.
The difference lies in the source, not necessarily in the physics of propagation.
15.98 Continuous run
The network must be able to run:
indefinitely
without an implicit:
reset after sample.
This is fundamental for the continuity of the percept.
15.99 Trial boundaries are analytical, not physical
An experiment may have:
trial 1
trial 2.
But if we are testing continuity, a trial boundary must not automatically reset the neuronal state.
A reset must be an explicit intervention.
15.100 Blank interval
The framework must explicitly support:
no sensory input
with the network at the same time:
network continues to evolve.
15.101 Sleep-like offline period
For Deep State Learning a regime may later be introduced:
sensory disconnected
spontaneous dynamics active
selected plasticity active.
This may be used for experiments with consolidation.
15.102 State freeze
For some control experiments it will be useful to:
freeze plasticity
or:
freeze oscillator phase
or:
freeze synaptic weights.
In this way the dynamic components can be separated.
15.103 Replay of the sensory input
It must be possible to:
record
the sensory stream and then:
replay exactly.
Two networks will then receive an identical external world.
15.104 Replay of internal events
For some control experiments it may be useful also to replay:
exact internal spike sequence.
This will make it possible to distinguish:
architecture response
from:
stochastic event generation.
15.105 State comparison
The framework must contain the functions:
compare_states(S_A, S_B)
compare_trajectories(T_A, T_B).
The resulting metric may be pluggable.
15.106 Trajectory distance
Possible metrics:
Euclidean latent distance,
dynamic time warping,
cosine similarity,
classifier separability.
It is not necessary to encode one method into the theory.
15.107 Transition detection
The framework must detect:
M_A -> M_B
on the basis of an external analytical model.
This will make it possible to measure:
dwell time,
transition probability,
hysteresis.
15.108 Transition matrix
After several runs we create:
P_ij.
This is a key object for the analysis of the Perceptual Manifold.
15.109 Deep Percept metrics
The framework must be able to compute or export the basis for:
decodability,
persistence,
metastability,
history dependence,
temporal sensitivity,
prediction relevance,
causal impact,
robustness,
generalization,
integration.
15.110 No single consciousness score
Cognia must not have:
consciousness = 0.83.
That would be methodologically misleading.
There may exist:
DeepPerceptMetrics
as a vector of functional properties.
15.111 Debug mode
A research engine must allow detailed debugging:
why did neuron fire?
which spikes arrived?
what was phase?
what changed weight?
Without this, the interpretation of small experiments will be difficult.
15.112 Visualizer
A visualization tool will later be very useful:
raster plot,
population activity,
oscillator phase,
state-space trajectory,
transition graph.
It is not necessary for the function of the network itself, but it will help the research fundamentally.
15.113 Deterministic reproduction with a fixed seed
For the same:
network,
input,
seed,
engine version
the following must hold:
exact event log reproducible
within a single numerical environment.
This is a basic validation condition.
15.114 Numerical precision
The engine must explicitly define:
time precision.
For example:
nanoseconds,
microseconds,
floating-point seconds.
There must be no random reordering of events because of numerical rounding.
15.115 Temporal granularity is not a neuronal clock
If the engine uses, for example:
1 µs resolution,
this is not a:
neural update clock.
It is merely the numerical precision of the representation of a timestamp.
15.116 Parallel execution
A later optimization may process independent events in parallel.
It must, however, preserve:
causal ordering.
A performance optimization must not change the physics of the experiment.
15.117 GPU execution
Cognia may later make use of the GPU.
But batch processing must not inadvertently introduce a:
global synchronous update.
If the GPU requires batching, it must be clearly distinguished:
implementation batching
from:
model synchronization.
15.118 Validation suite
Before the DPSH experiments there must exist unit/integration tests for:
event ordering,
delay,
refractory period,
stochastic probability,
phase calculation,
STDP timing,
snapshot/restore,
random replay.
15.119 Test event ordering
For example:
A at 10 ms
B at 9 ms.
The queue must always deliver:
B before A.
15.120 Test non-commutative ordering
Create a neuron for which:
A -> B
leads to:
state X
and:
B -> A
to:
state Y.
The engine must preserve this difference.
15.121 Test delay
A spike:
source at 10 ms
delay 5 ms
must arrive at:
exactly 15 ms.
15.122 Test random reproducibility
The seed:
42
must generate the same sequence of stochastic events.
15.123 Test snapshot/restore
After a:
snapshot at t=100 ms
the restored run with the same seed must create the same future trajectory, provided no intervention is carried out.
15.124 Test branch intervention
From one snapshot:
branch A unchanged
branch B phase shifted.
The difference in trajectory can then genuinely be attributed to the intervention.
15.125 Test STDP
Define:
pre at 10 ms
post at 15 ms.
Verify the expected:
Δw.
Then reverse the order.
15.126 Test oscillator
Verify:
phase(t)
and the modulation of the neuronal response.
15.127 Minimal DPSH engine 0.1
For the first experiments it is not necessary to implement everything.
The minimal version needs:
- an event-driven scheduler,
- simulation time,
- a stateful neuron,
- stochastic firing,
- a refractory period,
- excitatory/inhibitory synapses,
- delays,
- an explicit local oscillator,
- oscillator modulation,
- recurrence,
- fixed-seed randomness,
- a spike/state recorder,
- snapshot/restore,
- phase scramble,
- mechanism ablation.
Plasticity may come in the next iteration if the first experiment tests only the dynamics.
15.128 Cognia DPSH engine 0.2
The next version will add:
STDP,
plasticity logging,
learning gates,
spontaneous replay,
manifold comparison.
15.129 Cognia DPSH engine 0.3
After that:
prediction,
prediction-error modulation,
multiple sensory channels,
precision.
15.130 Cognia DPSH engine 0.4
Subsequently:
Global Workspace,
cross-module broadcast,
top-down feedback.
15.131 The first experimental architecture
The first network should be very small.
For example:
Population A
Population B
with:
recurrent excitation within population,
mutual inhibition,
stochastic baseline,
local oscillatory modulation.
The task is only to find out:
can metastable symmetry-broken states emerge?
15.132 The second architecture
Extend by:
sensory A,
sensory B,
ambiguous X.
The test:
A -> blank -> X
B -> blank -> X.
This will be the first Deep Percept context experiment.
15.133 The third architecture
Add:
phase manipulation.
The test:
intact phase
vs
scrambled phase.
At a matched firing rate.
15.134 The fourth architecture
Add:
STDP.
To test:
experience changes state-space geometry.
15.135 The fifth architecture
Add:
predictive continuation.
For example:
A -> B -> C sequence.
15.136 The sixth architecture
Add:
second modality
and:
shared downstream modules.
To test:
Perceptual Manifold integration.
15.137 The seventh architecture
Only after that:
Global Workspace.
This will prevent too many mechanisms from being mixed together in one experiment.
15.138 Minimal data format of an experiment
Every run should generate:
metadata.json
events.csv
states.csv
synapses.csv
metrics.json.
For example:
metadata:
theory_version
engine_version
experiment
seed
parameters.
15.139 Standard experimental report
Every experiment should end with a structured report:
Hypothesis
H3.
Manipulation
phase scramble.
Controls
matched rate,
matched spike count.
Primary metric
state separability.
Secondary metrics
dwell time,
transition entropy.
Result
...
Falsification status
supported / weakened / falsified / inconclusive.
15.140 Inconclusive is a valid result
If an experiment does not distinguish between mechanisms, the result must be:
inconclusive.
It must not be automatically interpreted as support.
15.141 Separation of development and experiment
Once an experiment has been defined, a change of the engine during an experimental series must create a new:
engine version
and the experiment must be run again.
Otherwise the results cannot be compared.
15.142 Architectural principles of the Cognia language
From the point of view of the DSL, the basic constructs should approximately be:
neuron
population
synapse
oscillator
memory
modulator
controller
input
workspace
plasticity
experiment.
Each must have clearly separated semantics.
15.143 The event as first class
For DPSH it may be very useful to introduce:
event
as an explicit first-class concept of the language or the runtime.
Because a large part of the theory concerns:
timing,
ordering,
propagation.
15.144 Delay as first class
Similarly:
delay
must not be merely an internal technical detail of the synapse.
It is an experimental variable.
15.145 Phase as first class
The oscillatory:
phase
must be available to:
modulation rules,
plasticity rules,
recorder,
experiment interventions.
15.146 Internal state as first class
A neuron or another dynamic construct must have:
state { ... }.
This will make it possible to create biologically as well as abstractly inspired units.
15.147 Mechanism composition
Cognia should support the composition of mechanisms.
For example:
stochastic neuron
+
oscillator modulation
+
STDP synapses.
Not the creation of a new hardcoded type for every combination.
15.148 Experimental transparency
Every mechanism must be visible in the source description.
A hidden automatic optimization of the engine that changes:
timing,
firing,
connectivity
may invalidate the experiment.
15.149 No implicit learning
The engine must not change weights unless a:
plasticity rule.
is explicitly active.
This is necessary for the controls.
15.150 No implicit reset
The engine must not automatically reset:
neuron states,
oscillator phases,
random streams
between stimuli.
15.151 No implicit synchronization
The engine must not implicitly align local events into a single tick unless that is explicitly an experimental control condition.
15.152 The main implementation principle
The Cognia DPSH engine must respect:
The global state must not be directly constructed by the engine. The engine provides only local dynamic mechanisms and precise temporal infrastructure. Macroscopic perceptual states must arise as the result of interactions defined by the neuronal architecture.
15.153 The main experimental principle
Every claim must take the form:
mechanism
->
predicted observable.
Then:
intervention
->
predicted change.
If the change does not occur:
hypothesis weakened.
15.154 Primary requirements for the nearest implementation
For the first real development phase of Cognia, DPSH recommends the following priority:
P0 – essential
event queue
absolute simulation timestamps
synaptic delays
stateful autonomous neuron
stochastic baseline firing
refractory state
local oscillator
phase modulation
recurrence
excitatory/inhibitory connections
snapshot/restore
fixed seed
event/state logging.
P1 – first experiments
phase scrambling
timing scrambling
rate-matched controls
population recorder
trajectory export
macrostate analysis.
P2 – learning
STDP
learning gate
synaptic logging
frozen replay
spontaneous-learning condition.
P3 – higher architecture
predictive loops
multi-modal inputs
shared manifold projections
Global Workspace.
15.155 Chapter research hypothesis
We formulate the technical hypothesis H13:
H13 – Experimental Realizability Hypothesis
If the key mechanisms of DPSH are genuinely causally relevant, it must be possible to realize them as separable local components of an event-driven neuronal system, and their presence or removal must create reproducible and specific changes in the state-space dynamics without the necessity of explicitly constructing a global perceptual state in the engine.
The stronger technical prediction:
A minimal Cognia network formed by autonomous stochastic neurons, recurrent excitatory and inhibitory connections, synaptic delays and local oscillatory modulation must be able to create measurable metastable macrostates without an explicit memory register or a central percept selector.
If even such a minimal architecture does not create the assumed dynamics, the corresponding part of DPSH must be revised before further higher mechanisms are added to the system.
15.156 The immediate implementation goal
The first implementation milestone is therefore not:
create conscious Cognia.
It is:
To create a reproducible event-driven experimental engine in which a run can be branched from exactly the same initial state with one controlled intervention, and the resulting neuronal trajectories can subsequently be quantitatively compared.
Once this framework exists, it is possible to begin genuinely testing DPSH.
From this moment on, the main task ceases to be the further extension of the theory.
The main task is:
implement
->
measure
->
falsify
->
revise.