Circadian Predictive Coding is a predictive-coding-based learner with sleep-phase structural plasticity. It tracks per-neuron chemical usage, modulates plasticity during wake, and applies split/prune consolidation during sleep.
- Research and educational experimentation with biologically inspired learning dynamics.
- Controlled benchmarks against backpropagation and traditional predictive coding.
- Safety-critical production decisions.
- Unreviewed deployment in medical, legal, or financial decision pipelines.
- Base: predictive coding with iterative hidden-state inference.
- Extension: circadian dynamics:
- chemical accumulation and decay
- plasticity gating
- reward-modulated wake learning (optional)
- adaptive sleep triggers
- adaptive sleep budget scaling (optional)
- structural split/prune
- optional rollback and homeostatic controls
- Toy two-cluster synthetic dataset (NumPy experiments).
- Synthetic and torchvision-backed vision datasets (ResNet benchmark workflow), including CIFAR-10/CIFAR-100.
Primary comparison metrics:
- Test accuracy
- Cross-entropy / energy
- Training throughput (
samples/s) - Inference latency (
mean,p95) and throughput - Circadian adaptation telemetry (splits, prunes, hidden dimension trajectory, rollbacks)
- Benchmark conclusions are sensitive to sleep hyperparameters.
- Circadian adaptation can underperform if split/prune schedules are too aggressive.
- Current implementation focuses on head-level circadian adaptation in ResNet benchmarks.
- No personal data is required by default benchmark workflows.
- Public benchmark claims should include dataset, seeds, and configuration details for reproducibility.
Active research repository; APIs and defaults may evolve. Use release tags for stable references in external projects.