A Scalable Hardware-Efficient CSNN Accelerator with Binarized STDP and STDG-based Hybrid Learning

Abstract

Spiking neural networks (SNNs) offer a biologically inspired, energy-efficient computing paradigm that processes information through sparse, event-driven spike signals. In this work, we present Binarized spike-timing dependent plasticity-gradient (BSTDPG), a novel SNN framework that combines Time-to-first-spike (TTFS) latency encoding with simple integrate-and-fire (IF) neuron model, binarized convolutional layers with local Spike-timing dependent plasticity (STDP) learning rule, and a gradient-based classifier. To support hybrid learning, the final output layer is trained using either a Support Vector Machine (SVM) or gradient-based backpropagation, allowing flexible deployment in both unsupervised and supervised settings.

The proposed architecture consists of two spiking convolutional layers followed by maxpooling and a classifier. A key innovation is the zero-skip optimization technique, which dynamically prunes inactive spike regions based on bounding boxes generated from TTFS-encoded spike trains. This drastically reduces unnecessary computations during training and inference. For instance, processing sample digit '0' during its first timestep without optimization results in full 28 × 28 convolution and 7860 ns latency. With zero-skip enabled, the latency drops to 1710 ns, achieving a 4.6× speedup and 78.2% processing time reduction. A similar improvement is observed at the full network level, where zero-skip reduces prediction latency from 1459 µs to 1212 µs, saving 246 µs. We evaluate BSTDPG on Modified National Institute of Standards and Technology database (MNIST) and Fashion-MNIST (F-MNIST) across both shallow and deep network configurations. The model achieves accuracies of 99.1% and 91.0%, respectively, outperforming or matching several state-of-the-art SNNs such as BANN, BS4NN, S4NN, and even deep temporal SNNs, despite using fewer layers.

Together, these innovations highlight BSTDPG as a promising architecture for low-latency, low-power neuromorphic applications including edge-AI, robotics, and event-based vision, offering an effective trade-off between biological plausibility, efficiency, and learning performance.