Enabling Efficient Inference of Convolutional Neural Networks via Approximation

Abstract

With the rapid advancements in the Artificial Intelligence area, Neural Networks (NNs) became the driving force both in general purpose and embedded computing domains. Especially, resource constrained embedded systems progressively rely on multiple NNs to provide on the spot sophisticated services. Nevertheless, supporting NN-based workloads is challenging due to the enormous computational and energy requirements. Exploiting the inherent error resiliency of NNs, significant research focuses on designing approximate Convolutional NN (CNN) inference accelerators, demonstrating that, for negligible accuracy loss, they satisfy tight latency, power, and temperature constraints. This chapter provides a comprehensive discussion of different aspects of approximate CNN implementations.