Compression, generation, and inference via supervised learning
Abstract/Contents
- Abstract
- Artificial intelligence and machine learning methods have seen tremendous advances in the past decade, thanks to deep neural networks. Supervised learning methods enables neural networks to effectively approximate low-level functions of human intelligence, such as identifying an object within an image. However, many complex functions of human intelligence are difficult to solve with supervised learning directly: humans can build concise representations of the world (compression), generate works of art based on creative imaginations (generation), and infer how others will act from personal experiences (inference). In this dissertation, we focus on machine learning approaches that reduce these complex functions of human intelligence into simpler ones that can be readily solved with supervised learning and thus enabling us to leverage the developments in deep learning. This dissertation comprises of three parts, namely compression, generation, and inference. The first part discusses how we can apply supervised learning to unsupervised representation learning. We develop algorithms that can learn informative representations from large unlabeled datasets while protecting certain sensitive attributes. The second part extends these ideas to learning high-dimensional probabilistic models of unlabeled data. Combined with the insights from the first part, we introduce a generative model suitable for conditional generation under limited supervision. In the third and final part, we present two applications of supervised learning in probabilistic inference methods: (a) optimizing for efficient Bayesian inference algorithms and (b) inferring the agents' intent under complex, multi-agent environments. These contributions enable machines to overcome existing limitations of supervised learning in real-world compression, generation, and inference problems.
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 | 2021; ©2021 |
| Publication date | 2021; 2021 |
| Issuance | monographic |
| Language | English |
Creators/Contributors
| Author | Song, Jiaming |
|---|---|
| Degree supervisor | Ermon, Stefano |
| Thesis advisor | Ermon, Stefano |
| Thesis advisor | Ma, Tengyu |
| Thesis advisor | Sadigh, Dorsa |
| Thesis advisor | Wu, Jiajun, (Computer scientist) |
| Degree committee member | Ma, Tengyu |
| Degree committee member | Sadigh, Dorsa |
| Degree committee member | Wu, Jiajun, (Computer scientist) |
| Associated with | Stanford University, Computer Science Department |
Subjects
| Genre | Theses |
|---|---|
| Genre | Text |
Bibliographic information
| Statement of responsibility | Jiaming Song. |
|---|---|
| Note | Submitted to the Computer Science Department. |
| Thesis | Thesis Ph.D. Stanford University 2021. |
| Location | https://purl.stanford.edu/hc981qt9024 |
Access conditions
- Copyright
- © 2021 by Jiaming Song
- 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...