Efficient Implementation of a Multi-Layer Gradient-Free Online-Trainable Spiking Neural Network on FPGA

• DOI: 10.48550/arxiv.2305.19468
• Publication Date: 2023-01-01
• Authors: Ali Mehrabi; Yeshwanth Bethi; André van Schaik; Andrew Wabnitz; Saeed Afshar
• Keywords: FOS: Computer and information sciences; field programmable gate array (FPGA); Artificial Intelligence (cs.AI); Spiking neural networks; Computer Science - Artificial Intelligence; Hardware Architecture (cs.AR); neuromorphic hardware; Computer Science - Neural and Evolutionary Computing; Electrical engineering. Electronics. Nuclear engineering; Neural and Evolutionary Computing (cs.NE); Computer Science - Hardware Architecture; supervised learning; TK1-9971

This paper presents an efficient hardware implementation of the recently proposed Optimized Deep Event-driven Spiking Neural Network Architecture (ODESA). ODESA is first network to have end-to-end multi-layer online local supervised training without using gradients and has combined adaptation weights thresholds in hierarchical structure. research shows that architecture can be implemented efficiently on a large scale hardware. The consists (SNN) individual modules for each layer enable self-learning back-propagation. By simple adaptive selection thresholds, Winner-Takes-All (WTA) constraint layer, modified weight update rule more amenable hardware, trainer module allocates neuronal resources optimally at having pass high-precision error measurements across layers. All elements system, including module, interact event-based binary spikes. hardware-optimized shown preserve performance original algorithm multiple spatial-temporal classification problems with significantly reduced requirements.