Probabilistic inference in the era of large models
Abstract/Contents
- Abstract
- Recent progression in generative artificial intelligence has witnessed a ballooning in model size and data dimensionality. These large models, however, come with increased computational demands which prohibit the use of many traditional probabilistic inference algorithms. There is a pressing need for new inference algorithms that are efficient enough to run on large models and modern architectures, and powerful enough to work with high dimensionalities and large datasets. In this dissertation, we address this challenge by designing algorithms using ingredients that are compatible with model scale, such as parallelization, amortized inference, and neural function approximations. We present a variety of techniques to improve sampling and inference, leading to faster sample speed, better flexibility of sample queries, and more accurate estimation of inference targets. Our methods are applicable to large models across a range of architectures such as diffusion, autoregressive, and masked-autoencoder models. We demonstrate these findings on applications spanning image and text generation, game-playing, and robotics domains. These insights lead to practical improvements and novel perspectives for efficiently deploying large-scale generative models.
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 | 2024; ©2024 |
Publication date | 2024; 2024 |
Issuance | monographic |
Language | English |
Creators/Contributors
Author | Shih, Bo Yun | |
---|---|---|
Degree supervisor | Ermon, Stefano | |
Degree supervisor | Sadigh, Dorsa | |
Thesis advisor | Ermon, Stefano | |
Thesis advisor | Sadigh, Dorsa | |
Thesis advisor | Ahmadipouranari, Nima | |
Degree committee member | Ahmadipouranari, Nima | |
Associated with | Stanford University, School of Engineering | |
Associated with | Stanford University, Computer Science Department |
Subjects
Genre | Theses |
---|---|
Genre | Text |
Bibliographic information
Statement of responsibility | Andy Shih. |
---|---|
Note | Submitted to the Computer Science Department. |
Thesis | Thesis Ph.D. Stanford University 2024. |
Location | https://purl.stanford.edu/nt653cd4468 |
Access conditions
- Copyright
- © 2024 by Bo Yun Shih
- 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...