GAQ-SNN: A Genetic Algorithm based Quantization Framework for Deep Spiking Neural Networks

The usage of Spiking Neural Networks (SNN) for edge-computing has become a major research topic over the years. However, a main challenge still remains, which is the high memory storage requirements of weights for large-scale SNN models. This could be a critical issues as the edge computing platform has tight constraints on the available on-chip memory. To address this issue, we proposed GAQ-SNN, a genetic algorithm based framework to reduce the requirements of memory weights while still maintaining good performance. This is accomplished via two major parts. Firstly, GAQ-SNN will find the optimal neural architecture for the SNN. Secondly, GAQ-SNN find the optimal quantization level for each layer of the SNN. Simulation and hardware implementation results show that, with GAQ-SNN, we could reduce the memory storage up to 12.5× while keeping the accuracy loss to 0.6% when compared to the baseline network.