Generating structures by editing prototypes
- Kelvin Guu.
- [Stanford, California] : [Stanford University], 2018.
- Copyright notice
- Physical description
- 1 online resource.
Also available at
At the library
All items must be viewed on site
Request items at least 2 days before you visit to allow retrieval from off-site storage. You can request at most 5 items per day.
|3781 2018 G||In-library use|
- Methods for structured prediction underlie many successful applications of machine learning, including machine translation, speech synthesis, image generation, protein structure prediction and many other problems. However, producing high-quality structures is challenging because the individual components of a structure (e.g., the words in a sentence or the pixels in an image) constrain and depend on each other in complex ways. A wide variety of approaches have been proposed to tackle this problem. At one extreme, retrieval-based methods sidestep the difficulty of modeling the internal consistency of structures by simply selecting from a pre-populated repository of ``good'' structures. However, this may sacrifice flexibility and coverage, as the repository could fail to contain every possible structure that might be required. At the other extreme, generation-based methods synthesize structures from scratch, often building them up one unit at a time. This offers extreme flexibility, but faces the difficult problem of modeling the complex dependencies that exist between units, as well as the challenge of searching over all possible structures for good configurations. We aim to combine the best of both worlds using a new approach called retrieve-then-edit: first, a structure is retrieved from a repository of high quality candidates (as in retrieval-based methods), and then it is edited into a new structure using a synthesis-based editor. The retrieval step ensures that we start in the neighborhood of an already high quality structure, while the editing step gives us the flexibility to customize the structure in a fine-grained way. We demonstrate that this approach can be successfully applied across diverse structured prediction problems spanning text generation, executable code generation, and reinforcement learning to interact with a web browser (where an agent's sequence of actions is treated as a structure).
- Publication date
- Copyright date
- Submitted to the Department of Statistics.
- Thesis Ph.D. Stanford University 2018.
Browse related items
Start at call number: