Automated Multisource Adaptation via Zero-shot INformation Generation (AMAZING)
Abstract
During this project, HRL Laboratories LLC, in collaboration with a world-class team from Johns Hopkins University (JHU, Prof. Philip Koehn) and the University of California Davis (UC Davis, Prof. Hamed Pirsiavash), made significant advancements in the fields of image classification, self-supervised learning, and machine translation, particularly in learning from limited data and adapting to novel environments. Their efforts under the Learn with Less Labels (LwLL) program were aimed at innovating systems that can learn from data even when labels are scarce or entirely absent. For self-supervised learning, they proposed a novel mean-shift algorithm that learns representations by grouping images together without contrasting between them or adopting much of prior on the structure or number of the clusters and created methods to incorporate constraints to the mean-shift algorithm, enabling us to leverage additional knowledge sources. They also developed robust defense mechanisms against sophisticated backdoor attacks tailored for transformer-based algorithms, thus bolstering the robustness of self-supervised learning.
Document Details
- Document Type
- Technical Report
- Publication Date
- Oct 01, 2023
- Accession Number
- AD1213132
Entities
People
- Amir Rahimi
Organizations
- Columbia University