Application-Tailored Artificial Intelligence (APTAI)
Abstract
The Application-Tailored Artificial Intelligence (APTAI) program will develop concepts for AI capabilities whose designs and learning processes are influenced both by training data and by key concepts and features proposed by experts in the intended application domains. In this third wave approach, learning processes will incorporate both human experience and patterns extracted from data to converge on informative features, representations, abstractions, and inductive strategies. One expected benefit of this approach is the emphasis on features that humans find salient, thus reducing the number of counterintuitive mistakes made by AI methods and minimizing the vulnerability to adversarial attacks, both in training data and in operations. Another expected benefit relates to human interaction, because the use of domain concepts provides a built-in framing for explanations to human partners. An additional expected benefit is the facilitation of more rapid learning processes including potentially transfer learning and one-shot learning. The APTAI program will make use of specific national security domains to drive assessment and transition of these initial concepts. Candidate domains include undersea autonomous navigation, imagery analysis, multi-modal reasoning, and games in support of strategy and planning.
Document Details
- Document Type
- Accomplishment
- Publication Date
- Oct 01, 2021
- Source ID
- 2a56d77f3a006285124725643f1e6854