Title: Disentangling Neural Networks Representations for Improved Generalization

Michael Cogswell
Ph.D. Candidate
School of Interactive Computing
Georgia Institute of Technology

Date: Tuesday, March 10th, 2020
Time: 12:00 pm to 2:00 pm (ET)
Location: Coda C1015 Vinings

BlueJeans: https://bluejeans.com/714664211

Committee:
Prof. Dhruv Batra, School of Interactive Computing, Georgia Institute of Technology
Prof. Devi Parikh, School of Interactive Computing, Georgia Institute of Technology
Prof. James Hays, School of Interactive Computing, Georgia Institute of Technology
Prof. Ashok Goel, School of Interactive Computing, Georgia Institute of Technology
Prof. Stefan Lee, Oregon State University

Abstract:

Despite the increasingly broad perceptual capabilities of neural networks, applying them to new tasks requires significant engineering effort in data collection and model design. In part, this is due to the entangled nature of the representations learned by these models. Entangled representations capture spurious patterns that are only useful for specific examples instead of factors of variation that explain the data generally. We show that encouraging representations to be disentangled makes them generalize better.

In this thesis we identify three kinds of entangled representations, enforce disentanglement in each case, and show that more general representations result. These perspectives treat disentanglement as statistical independence of features in image classification, language compositionality in goal driven dialog, and latent intention priors in visual dialog. By increasing the generality of neural networks through disentanglement we hope to reduce the effort required to apply neural networks to new tasks and highlight the role of inductive biases like disentanglement in neural network design.