Year: 2020
Type: dissertation
Abstract: In recent years, there has been an exponential rise in the quantity of data being acquired and generated. Machine learning provides a way to use and analyze this data to provide a range of insights and services. In this thesis, two popular machine learning algorithms are explored in detail and implemented in hardware to achieve high throughput and low latency. The first algorithm discussed is the Na¨ıve Online regularised Risk Minimization Algorithm. This is a Kernel Adaptive Filter capable of high throughput online learning. In this work, a hardware architecture known as braiding is proposed and implemented on a Field-Programmable Gate Array. The application of this braiding technique to the Na¨ıve Online regularised Risk Minimization Algorithm results in a very high throughput and low latency design. Neural networks have seen explosive growth in research in the recent decade. A portion of this research has been dedicated to lowering the computational cost of neural networks by using lower precision representations. The second method explored and implemented in this work is the unrolling of ternary neural networks. Ternary neural networks can have the same structure as any floating point neural network with the key difference being the weights of the network are restricted to -1,0 and 1. Under certain assumptions, this work demonstrates that these networks can be implemented very efficiently for inference by exploiting sparsity and common subexpressions. To demonstrate the effectiveness of this technique, it is applied to two different systems and two different datasets. The first is on the common benchmarking dataset CIFAR10 and the Amazon Web Services F1 platform, and the second is for real-time automatic modulation classification of radio frequency signals using the radio frequency system on chip ZCU111 development board. These implementations both demonstrate very high throughput and low latency compared with other published literature while maintaining very high accuracy. Together this work provides techniques for real-time inference and training on parallel hardware which can be used to implement a wide range of new applications.
Author: Stephen Tridgell
Institution: The University of Sydney
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W3048254329
