Towards Actionable Intelligence: Learning Representations for Decisions and Interventions
Abstract
Though we remain in the midst of an artificial intelligence renaissance, machine learning systems still have many critical weaknesses. They are highly sensitive to distribution shift, where small changes in how data is generated give rise to large drops in predictive performance. In turn, this means that machine learning cannot cope with dynamical data and cannot adjust to feedback from the physical world. ML systems are particularly vulnerable to strategic agents that interact with them. Further, ML systems suffer from significant diminishing returns, and memorizing remaining edge cases will require computational resources insurmountable for the department of defense. This project will address these shortcomings by determining new mathematical foundationsfor systems that use acquired data to enhance decision making in feedback with other complex, dynamic systems. We will study how to mitigate distribution shift using active feedback. We will investigate new, simpler representations for making decisions from complex, high dimensional perceptual data. We will aim to understand the fundamental limitations of the current machine learning paradigm where reactive feedback is computed by simple prediction systems. And we will explore how to incorporate rich perceptual information into feedback systems in robust, efficient, and reliable ways. If successful, this project will enable the unified analysis of the closed loop of learning, intervention, and manipulation. We aim to test these theoretical constructions and algorithms inrealistic settings in autonomous vehicles, robotics, social internet systems, and other applications of interest to the Office of Naval Research.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 08, 2020
- Source ID
- N000142012497
Entities
People
- Benjamin Recht
Organizations
- Office of Naval Research
- United States Navy
- University of California Regents