Doctoral Dissertations

Date of Award

5-2022

Degree Type

Dissertation

Degree Name

Doctor of Philosophy

Major

Computer Engineering

Major Professor

Husheng Li

Committee Members

Dali Wang, Wenjun Zhou, Jens Gregor

Abstract

Deep neural networks (DNNs) have achieved huge successes in various tasks such as object classification and detection, image synthesis, game-playing, and biological developmental system simulation. State-or-the-art performance on these tasks is usually achieved by designing deeper and wider DNNs with the cost of huge storage size and high computational complexity. However, the over-parameterization problem of DNNs constrains their deployment in resource-limited devices, such as drones and mobile phones.

With these concerns, many network compression approaches are developed, such as quantization, neural architecture search, network pruning, and knowledge distillation. These approaches reduce the sizes and computational costs of DNNs while maintaining their performance.

In this dissertation, we first focus on two of the most popular network compression schemes, i.e., network pruning and knowledge distillation. We aim to (1) develop more efficient network pruning approaches that can remove a large percentage of parameters/FLOPs from the DNNs while minimizing the performance degradation, and (2) train compact neural networks with the help of large, pre-trained networks under challenging scenarios in which limited information of the pre-trained networks are accessible. In the second part, we will develop efficient deep learning algorithms for a real-world application, i.e., modeling the biological cell migration process with deep reinforcement learning. The main contribution of this dissertation is summarized as follows.

We propose a novel network pruning approach, which removes filters based on the redundancy measurement in each layer. Different from existing works that prune the least important filters across all layers, we find that pruning filters from the layer with the most redundancy performs better.

We study knowledge distillation, which trains a compact network by mimicking the output of a pre-trained, over-parameterized network, under more challenging scenarios. In specific, we explore the possibility to learn from the pre-trained model when (1) the training set is not accessible and (2) the pre-trained model only returns top-1 index rather than probabilities.

We leverage efficient deep learning tools in the cell migration modeling with reinforcement learning, which helps reduce the training time. Therefore, novel biological mechanisms can be discovered within an acceptable period of time.

Download
Files over 3MB may be slow to open. For best results, right-click and select "save as..."

Share

COinS