H2Learn: High-Efficiency Learning Accelerator for High-Accuracy Spiking Neural Networks

Although spiking neural networks (SNNs) take benefits from the bioplausible modeling, low accuracy under common local synaptic plasticity learning rules limits their application in many practical tasks. Recently, an emerging SNN supervised algorithm inspired by backpropagation through time (BPTT) domain of artificial (ANNs) has successfully boosted SNNs, and helped improve practicability SNNs. However, current general-purpose processors suffer efficiency when performing BPTT for SNNs due to ANN-tailored optimization. On other hand, neuromorphic chips cannot support because they mainly adopt simplified implementation. In this work, we propose H2Learn, a novel architecture that can achieve high BPTT-based learning, which ensures At beginning, characterized behaviors learning. Benefited binary spike-based computation forward pass weight update, first design look-up table (LUT)-based processing elements engine update make accumulations implicit fuse computations multiple input points. Second, benefited rich sparsity backward pass, dual-sparsity-aware engine, exploits both output sparsity. Finally, apply pipeline optimization between different engines build end-to-end solution Compared with modern NVIDIA V100 GPU, H2Learn achieves $7.38\times $ area saving, notation="LaTeX">$5.74-10.20\times speedup, notation="LaTeX">$5.25-7.12\times energy saving on several benchmark datasets.