Learning to Learn and Adapt with Less Labels
Abstract
This document summarizes our research and technical contributions in the DARPA LwLL (Learning with Less Labels) program to reduce labeled data required in training and adapting machine learning models. Based on transfer learning, our approach leverages inductive biases acquired from various sources, including datasets, knowledge resources, models and architectures, existing tasks, and training algorithms. It integrates the aforementioned approaches into a coherent framework, aligned with the LwLL evaluation protocol, to learn in various learning conditions with limited supervision. Capitalizing on the recent advances in generative AI, this work studies novel and effective ways of injecting proper inductive biases to push the state of the art in image classification, object detection, video classification, and machine translation. The outcomes include extensive publications in top AI venues, workshop organization, and the training of postdocs and PhD students. The team has also excelled in LwLL evaluations, with the object detection system selected for NGA transition task. Furthermore, the knowledge and skills gained from the LwLL program, have been applied to our teams participation in the DARPA CCU program, where only 20 percent of the data is labeled.
Document Details
- Document Type
- Technical Report
- Publication Date
- Nov 07, 2023
- Accession Number
- AD1214497
Entities
People
- Chengyao Qian
- Ehsan Abbasnejad
- Gholamreza Haffari
- Hamid Rezatofighi
- Islam Nassar
- Mahsa Ghorbanali
- Mehrtash Harandi
- Samitha Herath
- Trang Vu
Organizations
- Monash University