Presentation description

Over the past decade, artificial neural networks (ANNs), especially deep neural networks (DNNs), have revolutionized the field of machine learning, making constant breakthroughs to achieve state-of-the-art results in various fields. However, their high computational and energy cost makes them less ideal for real-time autonomous systems that have to operate in resource-constrained environments, which have become a subject of interest in recent times. The demand for low-power, energy-efficient, and computationally cheap processing has caused the shift of focus to spiking neural networks (SNNs), which perform better in situations with less data and energy and are more biologically plausible.

Unlike ANNs, SNNs process information through discrete spikes, but the challenge in both neuroscience and computer science has been determining how to encode and decode that information. The dominant paradigm in neuroscience has been rate coding, where the rate of spikes fired determines meaning, but mounting evidence suggests that isn't the only explanation. An alternative theory is temporal coding, where information is encoded in the precise timing of spikes. This explanation aims to account for the fast-processing nature of our brains, which aligns well with the use cases of SNNs. Understanding how different temporal coding schemes affect the performance of SNNs will be crucial to making them viable for practical use on a larger scale.

Our research explores how temporal encoding and decoding schemes impact SNN performance, focusing on phase coding. We developed a spiking neural network from scratch using two layers of leaky-integrate-and-fire (LIF) neurons trained using an unsupervised spike-timing-dependent-plasticity (STDP) learning algorithm that adjusts weights based on the timing difference of pre-synaptic and post-synaptic spikes. We also implemented our own biologically inspired phase encoding and decoding scheme, which we compared to other temporal strategies using snnTorch, as well as a baseline ANN built with PyTorch, all evaluated on the Fashion-MNIST dataset.