New Paradigms for Edge Intelligence Hardware

点击翻译按钮获取中文摘要

The explosive growth of devices connected to the Internet, along with the arrival of the fifth generation wireless network, is leading to an exponential growth in the amount of data generated daily. Such data is typically sent to distant warehouse-scale cloud servers for processing with artificial intelligence (AI) algorithms such as deep neural networks. A number of factors are now contributing to the rise of edge intelligence, where such processing happens at or near where the data is collected. Our long-term goal is to develop technologies that enable pervasive AI to provide recommendations, decisions, and actions personalized to a local user and context. We will pursue this goal in the short-term by developing algorithms and hardware implementations for low-latency, low-power on-device training of deep neural networks and secure, ultra-reliable low-latency communications among edge devices and between edge devices and the cloud. Our short-term objectives are to (1) develop practical stochastic computing methods for on-device training of deep neural networks; (2) develop backpropagation-free on-device methods for training deep neural networks with invertible logic; and (3) implement ultra-reliable and low-latency decoders for short error-correcting codes. This aim of this proposed work is to enable the deployment of low-cost and energy-efficient intelligent devices providing pervasive AI to power emerging applications in personalized medicine, smart homes, Industry 4.0, robotics, and autonomous vehicles. As the proposed work lies at the intersection between algorithms and hardware, it provides an outstanding environment for training highly qualified personnel (HQP) to acquire knowledge and skills in developing advanced edge intelligence hardware and go on to make important contributions to Canadian industry and universities.