Deep convolutional neural networks and energyefficient hardware acceleration

Abstract

Deep Convolutional Neural Networks (CNN) have achieved state-of-the-art inference accuracy in a wide range of computer vision applications like image classification, object detection, semantic segmentation etc. Applications based on CNN require millions of multiply-accumulate operations to be performed between input pixels and kernel weights during inference. Such CNNs when realized on embedded devices or edge devices of the Internet of Things, a power/energy-efficient compute platform is required, which needs to meet the limited power/energy budget of the devices. There have been numerous works in literature to address these requirements. With a brief introduction to CNNs, state-of-the-art CNNs and hardware accelerators this talk will highlight on challenges faced in the accelerators, and solutions. In addition, this talk refers to a few prior works by the presenter, with topics of interest for further research.