Learning, memory and evaluation

Three different kinds of change

Cognia separates mechanisms that are often called “learning” together:

• neural state changes every step and carries short-term context;
• synaptic plasticity changes connection weights;
• structural plasticity adds or removes connections.

A fourth path, supervised tooling, trains declared output regions against labelled datasets. Always state which path produced a result.

Local plasticity

Hebbian and temporal rules strengthen correlated activity. Oja updates constrain the weight vector and usually preserve distributed inputs better than unconstrained Hebbian growth. Chemistry may scale the learning rate. Incoming-weight homeostasis limits runaway recurrent loops.

Omitting a plasticity rule freezes the edge. This is useful for fixed random projections and reservoirs.

Supervised readouts

cognia_train builds the model, measures it before training, trains and measures it afterwards. It can save the resulting network:

.\cmake-build-debug\cognia_train.exe model.cognia data.set 500 MyNetwork --softmax --save networks\model.bin

For classification, --softmax applies the delta learning_rate * pre * (target - softmax(logit)) to edges entering the output region. This is the recommended current path for fixed embeddings or reservoirs with a shallow classification head.

CHL is available experimentally, but deep credit assignment is not reliable without symmetric feedback weights. Do not present deep CHL results as equivalent to backpropagation.

Dataset modes

Independent datasets contain one sensory vector and one target per row. Evaluation resets neuron activity for each example.

Sequential datasets preserve state within each sequence. Use --sequential when order, recurrence, delays or polarization carry the information. Reset at sequence boundaries, not between tokens.

Memory mechanisms

Cognia can retain context through:

• neuron leak and recurrence;
• explicit delayed edges;
• adaptation and attractor-like dynamics;
• workspace ignition and focused settling;
• polarized module state;
• controller-driven operating regimes.

These mechanisms have different failure modes. Leak fades continuously, delayed edges retain a fixed horizon, workspace may monopolize attention, and polarization is a low-dimensional bias rather than an exact storage register.

How to evaluate a memory claim

Use an informative prefix followed by inputs that are identical across classes. Score the identical-input tail by depth. Compare with an ablation that differs only in the mechanism being tested. Test unseen sequence lengths, seeds and noise profiles.

Report at least:

• train/test separation;
• total and per-depth accuracy or regression error;
• confusion matrix or class coverage;
• number of seeds and confidence intervals;
• topology size and training command;
• negative results and transient settling errors.

Examples 06 and 07 demonstrate this style. Their results support specific retention properties, not production safety or general superiority over GRU, LSTM or state-space models.

Reproducibility

Version the model, generated datasets and generator. Set a seed. Keep an untrained or disabled-mechanism baseline. Save the trained .bin, but also preserve enough information to retrain it because the binary format is pre-1.0 and dimension-sensitive.