The Influence of Temporal Dependency on Training Algorithms for Spiking Neural Networks

In contrast to artificial neural networks (ANNs), spiking neural networks (SNNs), referred to as third-generation neural networks, exhibit behaviors that closely resemble those of the human brain. They are believed to offer more advantages in terms of energy efficiency and performance. Numerous training algorithms have surfaced for SNN, primarily falling into three main categories: ANN-to-SNN (ANN2SNN) algorithms, backpropagation through time (BPTT) algorithms, and spike-timing-dependent plasticity (STDP) algorithms. Among these, ANN2SNN algorithms lack temporal dependency, whereas BPTT and STDP algorithms do incorporate temporal aspects. In this study, we utilize two datasets to compare the performance of these three types of algorithms: the neuromorphic MNIST dataset with a time dimension and the conventional MNIST dataset without this temporal aspect. The comparison is conducted considering classification accuracy, dataset adaptability, and data dependency. The experimental findings indicate that the ANN2SNN algorithm is not compatible with the neuromorphic MNIST dataset. Moreover, the temporal-dependent BPTT and STDP algorithms notably outperform ANN2SNN in terms of data dependency. The connection between the temporal dependency of algorithms and their performance is further discussed. Moreover, the paper provides valuable insights into future research avenues and practical applications of SNNs.