Sample Efficient Learning for Deployment Time
Abstract
In scenarios envisaged in the context of ARO s autonomous platforms under its modernization initiative, AI agents encounter new environments, novel objects, and unforeseen or unscripted events. Black-box deep neural network (DNN) models that are at the forefront of data-driven technological solutions require large ground-truthed datasets, and are vulnerable to domain shifts and novel classes. As such collecting data in-situ for training is impossible, and we need sample efficient (small data) methods capable of rapidly adapting and learning to generalize from prior knowledge to these new situations is required. Our goal is to develop methods that minimize or even eliminate the need for annotated training data at deployment time. In particular, in this context of small data, our objective is to develop methods that seamlessly adapt to new environments that exhibit large domain shifts, and where new object classes manifest. We require that the resulting predictions of our method are nearly as good as a fully-trained DNN, which is trained on a sufficiently large ground-truthed dataset collected in the new environment. We outline a systematic approach that is based on leveraging data from well-sourced source domains and other side information to distill latent feature representations. While well-sourced tasks or domains could be leveraged, they are often significantly biased and weakly related to the deployed task, and directly transferring knowledge is a fundamental challenge. To address this challenge, we propose methods that learn signatures that are shared across different objects and environments. These representations allow for expressing an object into its constituent parts, and as such object/environments can be expressed with a shared vocabulary of parts. Nevertheless, objects are discriminable from each other, in terms of how they are put-together using this vocabulary. Finally, we propose novel scalable optimization methods based on regret minimization to learn these architectures and latent representations as well as efficiently adapt and generalize to new domains.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 25, 2021
- Source ID
- W911NF2110246
Entities
People
- Venkatesh Saligrama
Organizations
- Army Contracting Command
- Boston University
- United States Army