Memory and System Aware Architectures for Real-Time Machine Learning

In the last decade, there has been an explosion of growth in the field of Machine Learning (ML) enabled by the increased computing capacity of commodity devices. ML is being applied to an ever-increasing number of applications which require real-time operation placing an increased stress on current processing techniques. This dissertation consists of four works which target different aspects of the real-time ML problem.

The first work focuses on the task of performing k-Nearest-Neighbor (KNN) search on un-ordered point clouds for autonomous driving applications. Here, I introduce a compute architecture which accelerates kNN computation on an FPGA platform through alleviating memory bottlenecks and streamlining computation.

The remaining portions focus on the Augmented and Virtual Reality (AR/VR) platform as an emerging computing system. First, I introduce a lightweight ML-based compression algorithm which can reduce the energy consumption of high-speed video transmission from the AR/VR device. Next, I present an in-depth design space exploration of the multi-processor AR/VR system and consider how features of the ML algorithm impact the design of a near-sensor processor. Finally, I introduce a near-sensor processor architecture which can adapt to the large variation in workloads expected on a real-time AR/VR device.

Chair: Professor Zhengya Zhang