BrainAI Learnable Delays + Heterogeneous Time Constants

What problem does it solve?

BrainAI Learnable Delays + Heterogeneous Time Constants standardizes temporal expressivity in BrainAI's spiking neural networks by providing configurable, learnable delays (DCLS-style) and per-neuron time constants with flexible granularity, enabling ablations and thorough temporal analysis.

Core Features & Use Cases

• DCLS-style learnable delays with granularity options: per_synapse, per_output, per_input, per_block.
• Heterogeneous tau initialization and learnable per-neuron time constants for multi-timescale processing.
• Ablation matrix tooling to compare baseline, delays_only, hetero_only, and both across temporal benchmarks (SHD, sMNIST, SSC, MNIST).
• Diagnostics and checkpoints: state_dicts, metrics, and plots to quantify delay/tau behavior.

Quick Start

Enable delays and heterogeneous tau in BrainAI and run the four-config ablation on a temporal benchmark to observe impact on accuracy and timing.