RTSSNN: Efficient Image Classification for Latency-Critical and Energy-Constrained SNNs Through a Time-Step Reduction Technique

spiking neural networks (SNNs) are emerging as potent alternatives to convolutional neural networks (CNNs), especially for energy-constrained and latency-sensitive applications, due to their spiked activations and inherent sparsity. Rate-coded SNNs, trained with backpropagation through time (BPTT) and using static data over artificial time-steps, have achieved state-of-the-art results on benchmarks like Modified National Institute of Standards and Technology (MNIST). For resource-constrained devices, smaller models help meet memory and power limitations. However, shallow rate-coded SNNs often need numerous time-steps for accurate inference, increasing latency and computational cost. To address this, a technique named reduced time-step (RTS) is proposed to reduce time-steps, optimizing the balance between model size and convergence latency. RTSSNN leverages the periodic dynamics of leaky integrate-and-fire (LIF) neurons’ membrane potentials when stimulated with constant input by adding a small fully connected (FC) layer at the end of the network. At this depth, spikes are sparse and stable, allowing RTSs without losing accuracy. Demonstrated on four-bit quantized SNNs on Raspberry Pi, the method achieves 9x, 4x, and 4.9x operations reduction, 3x, 2x, and 2.2x time-step reduction, as well as 2.3x, 1.5x, and 1.9x runtime reduction during inference on MNIST, FashionMNIST, and german traffic sign recognition benchmark (GTSRB) datasets, respectively, with maintained accuracy. It also shows a 4x, 2x, and 1.9x operations, time-step, and runtime reduction in one-bit quantized SNNs. Additional experiments conducted on the higher complexity CIFAR-10 dataset as well as the dynamic neuromorphic N-MNIST dataset confirmed that RTSSNN is effective primarily on shallow networks with static datasets, where stable activations support periodic membrane behavior in LIF neurons.