S4NN

S4NN implements a supervised spike-based learning rule for training multilayer spiking neural networks to perform classification using rank-order temporal coding where each neuron emits a single spike per stimulus.

Key Features:

• Rank-order-coding: Encodes information in the temporal order of single spikes, with each neuron firing exactly one spike per stimulus.
• Spike-latency backpropagation: Uses a supervised learning rule analogous to error backpropagation that operates on spike latencies to compute approximated error gradients backward through the network.
• Readout-first-spike classification: Determines class by the first neuron to fire in the readout layer.
• Nonleaky integrate-and-fire neurons: Implements neurons using the nonleaky integrate-and-fire model.
• Multilayer fully connected SNNs: Applies to feedforward, multi-fully connected layer spiking neural networks regardless of depth.

Scientific Applications:

• Supervised image classification: Demonstrated performance with 97.4% test accuracy on MNIST and 99.2% on the Caltech Face/Motorbike dataset using multilayer fully connected SNNs.
• Temporal coding research: Supports studies of rank-order temporal coding and single-spike temporal encoding schemes.
• Neuron model evaluation: Enables comparison of nonleaky integrate-and-fire neuron models within complex network architectures.

Methodology:

Training applies a supervised spike-based rule that backpropagates approximated error gradients through feedforward multilayer fully connected SNNs by operating on spike latencies under rank-order-coding (one spike per neuron), with classification given by the first-spiking readout neuron and neurons modeled as nonleaky integrate-and-fire units.