Title: Learning from Augmentations: Data, Features, Interactions, and Knowledge

Date: Tuesday, April 18th

Time: 1:00 PM ET

Zoom link: https://gatech.zoom.us/j/4326036450

Chia-Wen Kuo

Robotics Ph.D. student

School of Electrical and Computer Engineering (ECE)

Georgia Institute of Technology

Committee:

• Dr. Zsolt Kira (Advisor)
• Dr. Chunyuan Li
• Dr. Dhruv Batra
• Dr. Judy Hoffman
• Prof. Larry Heck

Abstract:

The proposed research aims to improve deep learning model training by utilizing augmented training signals from data, features, and environments, as well as external knowledge to reduce the need for large-scale training datasets and model sizes. The research is divided into two main thrusts: (1) obtaining augmented training signals through data, features, and environments, and (2) augmenting models with external knowledge. The first thrust focuses on using data and feature augmentations to train models without relying on large labeled datasets, while also utilizing auxiliary training signals from the environment in vision and language navigation tasks. The second thrust proposes offloading knowledge to an external database and retrieving relevant knowledge for the target task, reducing the required model capacity for storing large-scale knowledge.