A Hardware-Aware Analysis of Energy Efficiency and Quantization Trade-Offs in Spiking Neural Networks
Spiking Neural Networks (SNNs) are one of the most promising bio-inspired neural networks models and have drawn increasing attention in recent years. The event-driven communication mechanism of SNNs allows for sparse and theoretically low-power operations on dedicated neuromorphic hardware. However, the binary nature of instantaneous spikes also leads to considerable information loss in SNNs, resulting in performance degradation. To address this issue, multi-level spiking neuron models have been proposed to provide both low-quantization error and minimal inference latency while approaching the performance of full precision Artificial Neural Networks (ANNs). In this paper, we provide an in depth analysis of the different trade-offs between quantization, generalization performance, and energy efficiency between binary SNNs, multi-level SNNs and ANNs for two different applications scenarios: image classification and image denoising. Our goal is to determine analytical thresholds on SNNs activity to understand when SNNs can be more energy efficient than ANNs but also to explain why SNNs cannot provide energy gains in other scenarios. Our analysis is based on an hardware execution model, which allows us to take into account the energy costs related to synaptic operations as well as memory accesses. Our experimental results show that multi-level spiking neurons provide better information compression, allowing therefore a reduction in latency without performance loss for classification tasks. When compared to binary SNNs on image classification scenarios, multi-level SNNs indeed allow reducing by 2 to 3 times the energy consumption depending on the number of quantization intervals. On neuromorphic data, multi-level SNNs allows us to drastically reduce the inference latency to 1 timestep, which corresponds to a compression factor of 10 compared to previously published results on binary SNNs. However, for the image denoising task we observe that neither the binary nor the multi-level SNNs can match the ANNs energy efficiency and performance. Based on a hardware-aware execution model of a neuromorphic SNN accelerator, we then provide analytical upper-bounds on SNNs activity and energy-efficiency. These theoretical bounds are then compared against the experimental results and allows us to understand, from a system level point of view, the relation between energy efficiency, quantization schemes and finally the performance of the SNNs.