Approximate Computing in Deep Learning System: Cross-Level Design and Methodology

Currently, deep learning is attracting massive researchers. As the scales of models and data are increasing rapidly, the application-specific acceleration systems are emerging. With inherent error resilience in machine learning systems, approximate computing, as a new computing paradigm, is widely adopted. The approximate computing is deployed in different design levels, including the algorithm, architecture, and circuit designs. Thus, the co-optimization and the co-design methods/architectures between different design levels are significant. This chapter intends to present the cross-level and systematic approximation techniques in deep learning processors, including designs and methodologies.