Flash: Fast Learning Via Auxiliary Signals, Structured Knowledge, and Human Expertise
Abstract
The original goal of the FLASH project was to develop novel and efficient machine learning algorithms that leverage rich forms of structured knowledge. Specifically, we build on the hypothesis that appropriate use of structured knowledge can substantially reduce the amount of hand-labeled data needed to achieve state-of-the-art performance on standard machine learning tasks, and address two key challenges: Leveraging structure: Develop general algorithms for leveraging structure to learn new concepts from few or no hand-labeled examples. Inferring structure: Develop general algorithms for inferring structure, either by actively learning from scratch or by transferring it from other domains The FLASH program executed this plan and produced a number of theoretical and practical contributions in all the areas mentioned above. In addition to developing theory, algorithms and representations we used those to develop applications in natural language and in computer vision.Moreover, in the course of the DARPA LwLL project, the focus of the field changed as a result of the success of large pre-trained generative AI models, including large language models (LLMs) like ChatGPT. While the goal of the work has not changed, our own research agenda adapted to these changes in the field while remaining close to the broad goals of our original proposal.
Document Details
- Document Type
- Technical Report
- Publication Date
- Nov 07, 2023
- Accession Number
- AD1214512
Entities
People
- Dan Roth
Organizations
- University of Pennsylvania