Strategy and Benchmark for Converting Deep Q-Networks to Event-Driven Spiking Neural Networks

Spiking neural networks (SNNs) have great potential for energy-efficient implementation of Deep Neural Networks (DNNs) on dedicated neuromorphic hardware. Recent studies demonstrated competitive performance SNNs compared with DNNs image classification tasks, including CIFAR-10 and ImageNet data. The present work focuses using in combination deep reinforcement learning ATARI games, which involves additional complexity as to classification. We review the theory converting extending conversion Q-Networks (DQNs). propose a robust representation firing rate reduce error during process. In addition, we introduce new metric evaluate process by comparing decisions made DQN SNN, respectively. also analyze how simulation time parameter normalization influence converted SNNs. achieve scores 17 top-performing Atari games. To best our knowledge, is first state-of-the-art multiple games Our serves benchmark DQNs paves way further research solving tasks