A low-power, high-accuracy with fully on-chip ternary weight hardware architecture for Deep Spiking Neural Networks

Recently, Deep Spiking Neural Network (DSNN) has emerged as a promising neuromorphic approach for various AI-based applications, such as image classification, speech recognition, robotic control etc. on edge computing platforms. However, the state-of-the-art offline training algorithms for DSNNs are facing two major challenges. Firstly, many timesteps are required to reach comparable accuracy with traditional frame-based DNNs algorithms. Secondly, extensive memory requirements for weight storage make it impossible to store all the weights on-chip for DSNNs with many layers. Thus the inference process requires continue access to expensive off-chip memory, ultimately leading to performance degradation in terms of throughput and power consumption. In this work, we propose a hardware-friendly training approach for DSNN that allows the weights to be constrained to ternary format, hence reducing the memory footprints and the energy consumption. Software simulations on MNIST and CIFAR10 datasets have shown that our training approach could reach an accuracy of 97% for MNIST (3-layer fully connected networks) and 89.71% for CIFAR10 (VGG16). To demonstrate the energy efficiency of our approach, we have proposed a neural processing module to implement our trained DSNN. When implemented as a fixed, 3-layers fully-connected system, the system has reached at energy efficiency of 74nJ/image with a classification accuracy of 97% for MNIST dataset. We have also considered a scalable design to support more complex network topologies when we integrate the neural processing module with a 3D Network-on-Chip.