Scaling human feedback
Abstract/Contents
- Abstract
- Human-generated data has been pivotal for significant advancements in artificial intelligence (AI). As AI models scale and are applied to a wider range of tasks, the demand for more and increasingly specialized human data will grow. However, current methods of acquiring human feedback, such as learning from demonstrations or preferences, and designing objective functions or prompts, are becoming unsustainable due to their high cost and the extensive effort or domain knowledge they require from users. We addresses this challenge by developing algorithms that reduce the cost and effort of providing human feedback. We leverage Foundation models to aid users in offering feedback. Users initially define their objectives (through language or a small dataset), and Foundation models expand this into more detailed feedback. A key contribution is an algorithm, based on a large language model, that allows users to cheaply define their objectives and train a reinforcement learning agent without needing to develop a complex reward function or provide extensive data. For situations where initial objectives are poorly defined or biased, we introduce an algorithm that efficiently queries humans for more information, reducing the number of needed queries. Finally, we conclude by proposing an information-gathering algorithm that eliminates the requirement for human intervention altogether, streamlining the feedback process. By making it cheaper for users to give feedback, either during training or when queried for more information, we hope to make learning from human feedback more scalable.
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 | Kwon, Minae |
|---|---|
| Degree supervisor | Sadigh, Dorsa |
| Thesis advisor | Sadigh, Dorsa |
| Thesis advisor | Goodman, Noah (Noah D.) |
| Thesis advisor | Yang, Diyi |
| Degree committee member | Goodman, Noah (Noah D.) |
| Degree committee member | Yang, Diyi |
| Associated with | Stanford University, School of Engineering |
| Associated with | Stanford University, Computer Science Department |
Subjects
| Genre | Theses |
|---|---|
| Genre | Text |
Bibliographic information
| Statement of responsibility | Minae Kwon. |
|---|---|
| Note | Submitted to the Computer Science Department. |
| Thesis | Thesis Ph.D. Stanford University 2023. |
| Location | https://purl.stanford.edu/sy876pv8068 |
Access conditions
- Copyright
- © 2023 by Minae Kwon
Also listed in
Loading usage metrics...