Advancing Artificial Intelligence for the Naval Domain
Abstract
The proposed research seeks to develop a principled integration of deep learning (DL) and knowledge representations (KR). The proposed program will address shortcomings of traditional DL, specifically the need for large annotated training sets, poor generalization beyond training data, brittleness under perturbation, and (often) shallow inference. By integrating DL and KR in aprincipled mathematical/statistical manner, we seek a new methodology that is capable of detecting and being robust to rare/unusual cases, as well as complex activities. To address the challenges of interest to the Navy, the proposed research seeks to realize agents that are capable of understandingand operating in uncertain, unstructured, open and dynamic environments, with the ability to predict events and plan actions.Integration of KR and DL presents several challenges, and necessitates new basic research. While domain knowledge is high-level and primarily symbolic, causal, relational, and semantic, learned representations and models are numerical and often distributed. For integration of knowledge and learning, good examples derive from insights from humans, who are typically good at domainknowledge adaptation. Therefore, while the proposed program focuses primarily on the methodological and principled integration of KR and DL, we also propose a cognitive-science component of the proposed program (from a budget perspective, this constitutes about 15% of the project, with the remaining over 85% directed toward mathematics, statistics and machine learning).We seek development of algorithms that may be applied to complex and heterogeneous data of interest to the Navy, allowing seamless integration of current knowledge and new experiences. To achieve this, we propose three principal areas of methodological research:??? DL models that are capable of learning with unlabeled data: We seek to integrate unsupervised DL with knowledge transfer, while also leveraging KR.??? Agent-based reinforcement learning (RL), integrating DL and KR: We propose a new DL methodology for RL data encoding, coupled with KR integration for both policy learning and execution.??? Leveraging and extensions of modern statistical natural language processing (NLP): Semantic representations are a natural means of constituting KR, and we will learn from NLP to encode other forms of KR. NLP is an important area for the Navy in its own right, and insights from NLP will also be leveraged to perform KR integration and knowledge transfer for modalities other than text (e.g., general imagery and video).The proposed research on DL and KR methodology will be supported by a fourth proposed area of focus,??? Leveraging insights from cognitive science: We will investigate knowledge transfer and learning in humans, for complex tasks, with human-subject experiments motivated by the above methodological directions.As an example of how these thrusts will be integrated, we will investigate methods for knowledge transfer of different types of (learned) task-optimized encoding strategies, as well as integration of such with KR. It is understood that humans appear to also manifest task-optimized encoding of data (e.g., text, images and video). Consequently, this is an interesting domain for investigating the relationships between algorithms and humans. The proposed team brings together expertise in statistical machine learning, deep learning, natural language processing, and cognitive science. Weekly team meetings and tight interdisciplinary research coupling are planned, leveraging the close proximity of the investigators. It is anticipated that the proposed research, while ambitious, is achievable within the proposed period of performance, and offers the potential to significantly advance the capability of the Navy to deploy DL-KR algorithms on increasingly autonomous missions.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 19, 2018
- Source ID
- N000141812871
Entities
People
- Lawrence Carin
Organizations
- Duke University
- Office of Naval Research
- United States Navy