A Survey of Approximation based Hardware Acceleration Techniques for Deep Neural Networks (Invited)

The rapid development of artificial intelligence (AI) requires substantial computing power based on massive hardware resources to support it. Thus, it poses a significant challenge for implementing AI systems including deep neural networks (DNNs). To improve the energy-efficiency, recent research has increasingly focused on approximate computing technology-based implementation for DNN acceleration. In this paper, approximate techniques from both the algorithm level and circuit level are introduced for DNN implementation. Besides, Error compensation and accuracy recovery methods for approximate DNNs are also demonstrated.