Spike Codec & Loss Pack

Batch-first spike encoding and decoding transforms inputs into a SpikeBatch of shape (B, T, N) while applying AMP-hardening to ensure stable gradient flow during spiking neural network training.

Spike-loss composition combines ProbSpikes, SpikeRateRegularization, TemporalConsistency, and MembraneRegularization via an SNNLossComposer to provide configurable differentiable losses with per-term diagnostics.

The spike codec and loss pack framework is explicitly AMP-safe, ensuring end-to-end gradient flow and differentiable loss composition remain stable under Automatic Mixed Precision configurations.

Combining temporal consistency and ISI regularization is best achieved through a modular loss composer that aggregates multiple spike-loss terms and provides per-term diagnostics for evaluation.

Reproducible spike encoding is managed through deterministic evaluation paths and explicit generator control, ensuring that SpikeBatch generation remains consistent across multiple experiments.

Gradient flow breaks during AMP stress testing when spike-loss terms lack AMP-hardening; using a dedicated codec with AMP-safe loss terms ensures end-to-end differentiability across configurations.