Generating Temporal Signals from Static Images for Spiking Neural Networks

Spiking neural networks (SNNs), i.e., neural architectures that represent and transmit information in the form of temporally distributed spikes, require time-dependent input, whereas data in computer vision are most commonly available as static images. This study addresses the transformation pipeline “image → temporal signal → spikes” and examines how the choice of input encoding influences SNN training dynamics, spike activity density, and computational cost. The experimental section implements and compares two encoding families: first-spike-time encoding (Latency) and intensity-based Poisson encoding (Poisson). Within these families, four operating modes are considered: baseline Latency without background suppression, modified Latency with a silence threshold, stochastic Poisson, and deterministic Poisson. The evaluation employs the following metrics: the average number of spikes per sample, the number of synaptic operations, an energy-related proxy metric, and indicators characterizing competition among hidden-layer neurons. Experiments conducted on the MNIST dataset (60 000 training and 10 000 test images) using a network with a hidden layer of 100 neurons and a simulation horizon of 200 time steps demonstrate that all examined modes support stable learning without activity collapse. Among them, the modified Latency mode with a silence threshold of x_min = 0.05 achieves the most favorable balance between useful activity and computational cost: at 323.41 spikes per sample, it requires 14925.09 synaptic operations, whereas the baseline Latency mode without background filtering, despite exhibiting a comparable level of output activity (311.22 spikes per sample), requires 78 400 synaptic operations.