Extracting and learning structures from data

In this work, we study models which explicitly capture and learn structures from data. For the task of supervised and unsupervised textual grounding, we propose a unified framework which links words to image concepts. A parameter between each word and image concept is learned and the learned parameters are easily interpretable. Next, for the task of generative modeling of multi-agent trajectories, we design models which share parameters based on the relationship between agents in the system to achieve permutation equivariance. This representation is particularly suitable in a multi-agent setting where the identity of the agents is unknown. We achieve better performance than conventional fully connected deep nets. Lastly, we present a framework on how to learn equivariance properties from data; this framework is based on learning how to share parameters in a model. We provide analysis on Gaussian vectors in terms on mean squared error criterion and empirically show that our approach can recover shift and permutation equivariances.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Electrical & Computer Engr

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2021

Author dc:creator

• Yeh, Raymond A

Contributors dc:contributor

• Schwing, Alexander G
• Hasegawa-Johnson, Mark
• Do, Minh N
• Forsyth, David

• machine learning
• computer vision

Statement dc:rights

• Copyright 2021 Raymond Yeh

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/110702

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/110702