Sample-efficient learning with self-supervision

Zhong, Yuanyi

Permalink

https://hdl.handle.net/2142/117813 Copy

Description

Title

Sample-efficient learning with self-supervision

Author(s)

Zhong, Yuanyi

Issue Date

2022-12-01

Director of Research (if dissertation) or Advisor (if thesis)

• Wang, Yuxiong
• Peng, Jian

Doctoral Committee Chair(s)

Wang, Yuxiong

Committee Member(s)

• Forsyth, David
• Zhang, Lei

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• deep learning
• self-supervised learning
• computer vision
• reinforcement learning
• sample efficiency
• artificial intelligence

Abstract

Sample efficiency is a crucial aspect of machine learning algorithms. A more sample-efficient algorithm can achieve better performance with less training data. On the other hand, the success of modern machine learning relies heavily on large data samples. For example, a visual recognition system typically needs to be trained on thousands or millions of labeled images to reach good accuracy. In addition, a reinforcement learning agent needs to experience a vast number of environmental interactions to be competitive. With the ever-growing demand for real-world machine learning solutions, it becomes imperative to develop more sample-efficient algorithms that rely less on human annotations or environment interactions and thus can learn more efficiently. Learning from self-supervision has recently emerged as a promising way to improve sample efficiency. The idea is that we can train the model with supervision signals generated not by human annotators or the environment but rather by the carefully constructed self-supervised tasks based on the model itself. In this dissertation, I present my contributions to improving the sample efficiency of machine learning with self-supervision for visual and reinforcement learning problems. The dissertation naturally breaks down into three parts according to the techniques used. The first two parts contain methods designed for vision tasks such as object detection and image segmentation. The core idea is either to leverage pseudo labels generated by an existing model, as in Chapters 2-3, or to leverage the invariance and equivariance property that we know to be inherently true for natural images, as in Chapters 4-6. The third part describes my research projects on building more sample-efficient reinforcement learning algorithms based on the ideas of self-supervised video prediction, self-play against a population, and self-critic with variance reduction as in Chapters 7-9. Altogether, this dissertation demonstrates multiple effective ways to improve the sample efficiency of machine learning across a wide range of tasks and the tremendous potential of self-supervision in the development of future AI systems.

Graduation Semester

2022-12

Type of Resource

Thesis

Copyright and License Information

Copyright 2022 Yuanyi Zhong

Owning Collections