Expand description
neuBurn — Spiking neural network framework on top of Burn. Rust-native alternative to snntorch.
Modules§
- bntt
- BatchNorm through time (BNTT): separate BatchNorm parameters per time step. Used in SNNs to normalize activations at each time step independently.
- encoding
- Spike encoding utilities: rate, latency, delta.
- event_
driven - Event-driven SKAN: delta-gated KAN + LIF layers.
- functional
- Loss functions and utilities for SNN training.
- kan
- Kolmogorov-Arnold Network (KAN) layer.
- layers
- Composite SNN layers: Conv+LIF (SpikingConv2d), etc.
- neuron
- Spiking neuron modules (LIF, Synaptic, Alpha, RLeaky, Lapicque).
- prelude
- Prelude: most commonly used types.
- slstm
- Spiking LSTM: LSTM whose output is passed through a spike (surrogate) nonlinearity. Input [batch, seq_len, input_size] -> spiked output [batch, seq_len, hidden_size].
- state
- Neuron and layer state types (membrane potential, synaptic current, trace). State is not part of Module/Record; it is passed in/out of step(). ResetMode controls how membrane potential is reset after a spike (snntorch-style).
- surrogate
- Surrogate gradients for backpropagation through spike (Heaviside) nonlinearity.