Stochastic Spiking Neural Networks with First-to-Spike Coding
Yi Jiang, Sen Lu, Abhronil Sengupta
TL;DR
This work addresses the inefficiency of rate-coded deterministic SNNs by integrating stochastic neurons with First-to-Spike temporal coding and demonstrating scalable direct training on deep architectures. It presents direct-training frameworks for deterministic and stochastic SNNs with First-to-Spike coding, including loss functions and surrogate-gradient/BPTT optimization, validated on MNIST, CIFAR-10, and architectures from 2-layer MLP to VGG. Key findings show First-to-Spike greatly reduces latency and can improve energy efficiency, while stochastic SNNs offer accuracy gains and improved noise robustness at some cost to sparsity and energy in certain regimes, providing a nuanced trade-off map for neuromorphic deployment. The work delivers a scalable, quantitatively analyzed approach to co-design stochastic temporally encoded SNNs, with implications for edge and neuromorphic hardware where fast inference and robustness are critical, and suggests directions toward larger datasets like ImageNet.
Abstract
Spiking Neural Networks (SNNs), recognized as the third generation of neural networks, are known for their bio-plausibility and energy efficiency, especially when implemented on neuromorphic hardware. However, the majority of existing studies on SNNs have concentrated on deterministic neurons with rate coding, a method that incurs substantial computational overhead due to lengthy information integration times and fails to fully harness the brain's probabilistic inference capabilities and temporal dynamics. In this work, we explore the merger of novel computing and information encoding schemes in SNN architectures where we integrate stochastic spiking neuron models with temporal coding techniques. Through extensive benchmarking with other deterministic SNNs and rate-based coding, we investigate the tradeoffs of our proposal in terms of accuracy, inference latency, spiking sparsity, energy consumption, and robustness. Our work is the first to extend the scalability of direct training approaches of stochastic SNNs with temporal encoding to VGG architectures and beyond-MNIST datasets.