Title: Efficient Visual Learning for Scene Understanding

Date: Tuesday, November 21, 2023

Time: 12:00 - 1:30 pm EST / 9:00 - 10:30 am PST

Location: https://gatech.zoom.us/j/7745230525

Yen-Cheng Liu

Machine Learning PhD Candidate

School of Electrical and Computer Engineering

Georgia Institute of Technology

Committee

1. Dr. Zsolt Kira (Advisor), School of Interactive Computing, Georgia Tech
2. Dr. Judy Hoffman, School of Interactive Computing, Georgia Tech
3. Dr. Larry Heck, School of Electrical and Computer Engineering and the School of Interactive Computing, Georgia Tech
4. Dr. Mark Davenport, School of Electrical and Computer Engineering, Georgia Tech
5. Dr. Diyi Yang, Computer Science Department, Stanford University

Abstract

Significant advancements in scene understanding have been driven by deep neural networks. These learning-based frameworks enhance performance through extensive training datasets and a large number of trainable parameters. However, they are less scalable and require substantial computational and financial resources. This dissertation investigates two aspects of efficient visual learning for scene understanding: label-efficient learning and parameter-efficient learning. To reduce label supervision in instance-level scene understanding tasks, we develop a series of semi-supervised learning frameworks. These frameworks improve the label efficiency under various detector architectures and unconstrained data settings. To reduce parameter usage in multi-task training, we re-evaluate parameter-efficient methods from NLP for scene understanding and then propose a more parameter-efficient method for vision architectures. These advancements demonstrate the practicality and adaptability of efficient learning frameworks in diverse, resource-constrained environments.