Advancing generative models for real-world applications
Abstract/Contents
- Abstract
- While generative models hold thrilling potential, their limited usability presents substantial challenges for their widespread adoption in real-world applications. Specifically, existing methods tend to amplify harmful societal biases ingrained in their training data and often fail to accurately reflect subjective user specifications such as style in the generated outputs. Furthermore, a notable performance gap exists when handling data distributions with unique structures, such as periodicity, restricting their applicability beyond image and text data. This dissertation considers all such facets to help construct safe, reliable generative AI systems for practical integration and deployment. First, we present a methodological framework for tackling the challenges of bias mitigation and controllability. Building on the classical approach of density ratio estimation (DRE), we develop techniques to correct a learned model distribution such that it exhibits characteristics that are more closely aligned with an alternative target distribution of interest. Together, these contributions not only yield a new theoretical framework for DRE, but also improve performance on a diverse set of downstream tasks such as domain adaptation, data augmentation, and mutual information estimation. Next, we present two real-world applications of these methods for societal applications. We demonstrate that: (a) our reweighted generative modeling framework successfully mitigates dataset bias, and (b) more controllable models can better customize AI-generated music to individual preferences and assist the creative process. Finally, we conclude by developing new learning algorithms that incorporate domain-specific inductive biases into generative models for wireless communications, as well as for discrete data distributions.
Description
| Type of resource | text |
|---|---|
| Form | electronic resource; remote; computer; online resource |
| Extent | 1 online resource. |
| Place | California |
| Place | [Stanford, California] |
| Publisher | [Stanford University] |
| Copyright date | 2023; ©2023 |
| Publication date | 2023; 2023 |
| Issuance | monographic |
| Language | English |
Creators/Contributors
| Author | Choi, Eun Young |
|---|---|
| Degree supervisor | Ermon, Stefano |
| Thesis advisor | Ermon, Stefano |
| Thesis advisor | Koyejo, Sanmi |
| Thesis advisor | Re, Chris |
| Degree committee member | Koyejo, Sanmi |
| Degree committee member | Re, Chris |
| Associated with | Stanford University, School of Engineering |
| Associated with | Stanford University, Computer Science Department |
Subjects
| Genre | Theses |
|---|---|
| Genre | Text |
Bibliographic information
| Statement of responsibility | Kristy Choi. |
|---|---|
| Note | Submitted to the Computer Science Department. |
| Thesis | Thesis Ph.D. Stanford University 2023. |
| Location | https://purl.stanford.edu/ky644hz0016 |
Access conditions
- Copyright
- © 2023 by Eun Young Choi
- License
- This work is licensed under a Creative Commons Attribution Non Commercial 3.0 Unported license (CC BY-NC).
Also listed in
Loading usage metrics...