BrainAI Spiking Core (LIF Family) + Surrogate Gradient Training

To prevent hidden state leakage in spiking neural networks, implement an explicit state contract for LIF-family neurons using reset_state, detach_state, and forward with carry_state, ensuring deterministic resets and clean sequence processing across CPU and GPU.

Surrogate gradient options available for training LIF neurons include ATan, FastSigmoid, and Straight-Through Estimator (STE), which stabilize gradient flow through non-differentiable spike generation functions during backpropagation.

Implement truncated BPTT for time-unrolled SNNs by using the unified snn_unroll utility, which processes sequential data while explicitly managing neuron state carry and detachment to enable stable surrogate gradient training.

Yes, the explicit state management approach supports AdaptiveLIF and RecurrentLIF neurons, covering the full LIF-family to provide deterministic state resets, state carry, and robust debugging across various spiking neural network architectures.

Spiking neural network training destabilizes across sequence lengths due to hidden state leakage and unmanaged state carry, which can be resolved by applying an explicit state contract with deterministic resets and surrogate gradients.