- This event has passed.
Doctoral Dissertation Proposal Oral Presentation, Aaron Palmer
March 11 @ 1:00 pm - 2:00 pm EDT
Title: Optimization and Utility of Differentiable Goodness-of-Fit Statistics in Deep Learning
Ph.D. Candidate: Aaron Palmer
Major Advisors: Jinbo Bi
Associate Advisors: Zhiyi Chi, Derek Aguiar
Committee Members: Laurent Michel
Date/Time: Monday, March 11th, 2024, 1:00pm
Location: HBL Instruction 1102
Meeting link: https://uconn-cmr.webex.com/uconn-cmr/j.php?MTID=md43f59a84eb8b148135ec915e9c27720
Meeting number: 2630 622 3826
Password: 59qmQNQqan2
Abstract
Generative autoencoders have achieved impressive performance in representation learning and generative modeling tasks due, in part, to the augmentation of deep learning architectures with concepts from statistical theory and latent variable models. The training algorithm compresses and reconstructs model inputs while minimizing distortion and matching the encoding distribution on the latent space to a targeted prior; the latter condition is often seen as necessary for learning the data distribution to enable sample generation from a known distribution after model training. However, balancing the reconstruction and regularization loss components during model training with random samples is challenging and evaluating whether the encoding distribution is sufficiently proximal to a prior for high quality generation is typically implemented using heuristics or empirical evidence. In this proposal, we demonstrate how autoencoders benefit from goodness-of-fit hypothesis tests, which are statistical techniques that use random samples to decide if an unknown distribution is indistinguishable from that of a target distribution, and develop theoretically principled algorithms to optimize them. We propose leveraging Bayesian nonparametric priors and the statistical indistinguishability properties of the encoding distribution in our goodness-of-fit autoencoder to address the open set recognition problem, where the state of the environment is incomplete during training and testing may require evaluating novel classes.
