v 0.3.0
Sensory 0.3.0
Released August 4, 2026 · Windows x64 · 25.17 MiB
Local artifact: Sensory-0.3.0-windows-x64.zip
Sensory 0.3.0 introduces delayed sequence credit for Cognia inner layers. Eligibility traces allow a later next-token error to update plastic inner and recurrent edges that were active during preceding tokens, while preserving the exact softmax cross-entropy readout update.
Highlights
Three-factor eligibility traces
For a plastic inner edge i → j, the runtime maintains a decaying local trace:
e_ij(t) = 0.95 * e_ij(t-1) + pre_i(t) * phi'(post_j(t))
The output signal target - probability is carried up to three levels inward through the current graph. The weight update combines that learning signal, the temporal trace, the destination gain and the edge's declared learning rate. Fixed edges and the readout are excluded from this update.
The feature is opt-in:
cognia_train model.cognia data.seq 5 Network `
--sequential --softmax --eligibility
Training manifests accept the equivalent eligibility flag.
Streaming sequence trainer
The trainer no longer retains every input and target vector in memory. It indexes sequence-start file offsets, deterministically shuffles whole sequences, seeks each selected sequence and reads rows on demand. Compact indexed formats store one input token ID and one target ID per transition.
Inner controller training
Controllers can now train on stateful sequences, preserve activity until a blank-line boundary and update only their explicitly owned edges. Staged plans save and reload isolated controller checkpoints before the final readout stage.
New examples
- Example 08 establishes a safe numerical spam-filter baseline over UCI Spambase without distributing raw malicious email bodies.
- Example 09 demonstrates Czech dialogue, staged inner-controller training and repetition-cycle termination.
- Example 10 provides a lossless Czech syllable tokenizer with character fallback, sentence intent state, autoregressive generation and a FineWeb2 sample/filter/sequence pipeline.
Downloaded FineWeb2 documents, generated datasets and trained checkpoints are deliberately not included in the release ZIP. The example scripts create local input/ and output/ data when run.
Validation
- all 14 parser/GPU model tests pass;
- supervised classification remains 30/30;
- sequential controller training reports
changed-foreign=0; - eligibility regression changes all 12 plastic inner test edges;
- the Example 10 bootstrap changed 2,248 inner edges with total absolute delta
0.014303394; - REST firing and deterministic reset regressions pass;
- the full Sensory application and all Cognia command-line tools build successfully.
Research status and limits
Eligibility traces are an online e-prop-style approximation, not exact backpropagation through time. Credit decays over time, reverse transport is limited to three levels and training costs additional GPU edge traversals. The measured inner-edge changes prove that the mechanism is active; they do not yet prove lower held-out perplexity or better Czech generation.
Upgrade notes from 0.2.0
Eligibility learning is disabled unless explicitly requested, so existing sequence training retains its previous readout-only behavior. Saved .bin networks remain dimension- and version-sensitive. Retrain a model under 0.3.0 before comparing it with a newly eligibility-trained checkpoint.
Requirements
- 64-bit Windows 10 or 11;
- NVIDIA GPU compatible with the included Compute Capability 7.5 build;
- current NVIDIA driver supporting CUDA 13.2.
SHA-256:
74237c61515326c55e6f8155f3d7d4e762b592d34a2ea210aafcca7964de96b0
Sensory 0.3.0 is experimental research software and is not safety-critical software.