Towards Streaming Perception and Action with Continual Adaptation

Abstract

(Approved for Public Release)This proposal addresses one of the fundamental challenges in real world deployment of machine learning systems, namely nonstationarity of the data and its detrimental effect on model training and test time performance. This can take the form of smooth changes that emerge over time, or abrupt discontinuities when the agent is switched to a new environment. We pro pose three streams of work that address both these sources. First, how to robustly learn representations that evolve along with the data distribution. Second, how to learn models of the underlying dynamics efficiently, enabling fast adaptation to new environments . Third, how do we incorporate inductive biases into training with the goal of robustifying the models against changes in the underl ying data distribution. Collectively, we anticipate that these methods will facilitate the deployment of Machine Learning and Reinfo rcement Learning in real world settings.

Document Details

Document Type
DoD Grant Award
Publication Date
Aug 20, 2021
Source ID
N000142112758

Entities

People

  • Lerrel Pinto

Organizations

  • New York University
  • Office of Naval Research
  • United States Navy

Tags

Fields of Study

  • Computer science

Readers

  • Adaptive Control and Estimation with Uncertainty in Dynamic Systems.
  • Agent-Based Social Robotics and Mobile-Assisted Learning in Virtual Environments.
  • Computational Fluid Dynamics (CFD)

Technology Areas

  • AI & ML
  • AI & ML - Autonomous Systems