Analysis of Input Samples and Internal Representations in Spiking Neural Networks from the Perspective of the Frequency Domain

Spiking neural networks (SNNs) have attracted the attention of researchers owing to their suitability for on-device intelligence. They have become a key technology for developing on-device intelligence products. However, SNNs can have difficulty achieving performance competitive with conventional ANNs, and improving their performance requires understanding how their internal representations are formed and change. Most existing studies focus on individual neurons or specific layers and do not provide a common metric for quantifying the frequency-domain differences between a reference representation and each hidden-layer representation. They also do not compute such a metric at each epoch to track how internal representations change during training. This limitation prevents researchers from understanding how SNNs' internal representations form and transmit the learned features. To address this limitation, we define layer-wise and model-level metrics for the frequency-domain differences between a reference representation and each hidden-layer representation and compute their values at each epoch to track how the differences change during training. We conduct experiments across various SNN architectures, datasets, and neuron configurations. Through our analysis, we discover that preserving the frequency distribution of input data samples in the internal representation is important for SNNs to process data properly.