LATERAL: Learning Automatic, Transfer-Enhanced, and Relation-Aware Labels
Abstract
Our team obtained excellent results in all of the official evaluations of the Darpa Learning with Less Labels (LwLL) program. We participated in the image classification, object detection, and machine translation tasks. For many of the checkpoints, we achieved the very best results across all performers, significantly outperforming the baseline provided by the JPL team. A key lesson learned to achieve these strong results was to focus on finding a good feature embedding, so that new tasks can be learned with just a few examples. In this vein, we proposed novel approaches for representation learning (across domains and modalities, using real and synthetic data), explored new model architectures, as well as transfer learning techniques. These research works have been published in top-tier AI conferences and several of them integrated into our high-performing systems delivered to Darpa for the official program evaluations.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jul 24, 2023
- Accession Number
- AD1206340
Entities
People
- Aadarsh Sahoo
- Assaf Arbelle
- Graeme Blackwood
- Joseph Schtok
- Leonid Karlinsky
- Rogerio Feris
Organizations
- International Business Machines Corporation (Armonk, NY)