Date of Award

Winter 12-22-2021

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Electrical Engineering and Computer Science

Advisor(s)

Qiu, Qinru

Subject Categories

Computer Engineering | Engineering

Abstract

Driven by the massive application of Internet of Things (IoT), embedded system and Cyber Physical System (CPS) etc., there is an increasing demand to apply machine intelligence on these power limited scenarios. Though deep learning has achieved impressive performance on various realistic and practical tasks such as anomaly detection, pattern recognition, machine vision etc., the ever-increasing computational complexity and model size of Deep Neural Networks (DNN) make it challenging to deploy them onto aforementioned scenarios where computation, memory and energy resource are all limited. Early studies show that biological systems' energy efficiency can be orders of magnitude higher than that of digital systems. Hence taking inspiration from biological systems, neuromorphic computing and Spiking Neural Network (SNN) have drawn attention as alternative solutions for energy-efficient machine intelligence.

Though believed promising, neuromorphic computing are hardly used for real world applications. A major problem is that the performance of SNN is limited compared with DNNs due to the lack of efficient training algorithm. In SNN, neuron's output is spike, which is represented by Dirac Delta function mathematically. Becauase of the non-differentiable nature of spike, gradient descent cannot be directly used to train SNN. Hence algorithm level innovation is desirable. Next, as an emerging computing paradigm, hardware and architecture level innovation is also required to support new algorithms and to explore the potential of neuromorphic computing.

In this work, we present a comprehensive algorithm-hardware codesign for neuromorphic computing. On the algorithm side, we address the training difficulty. We first derive a flexible SNN model that retains critical neural dynamics, and then develop algorithm to train SNN to learn temporal patterns. Next, we apply proposed algorithm to multivariate time series classification tasks to demonstrate its advantages. On hardware level, we develop a systematic solution on FPGA that is optimized for proposed SNN model to enable high performance inference. In addition, we also explore emerging devices, a memristor-based neuromorphic design is proposed. We carry out a neuron and synapse circuit which can replicate the important neural dynamics such as filter effect and adaptive threshold.

Access

Open Access

Download

Share

COinS