Leveraging Causal Structure for Prediction Across Environments

Abstract

In applications of machine learning ranging from autonomous driving to health care, data is often collected from a variety of settin gs (e.g., with varying sensors and geography), and machine learning is expected both to succeed at learning from this diverse data and to generalize to new settings unseen during learning. We formalize this problem through a causal lens, assuming that different "environments" share the same causal structure, but that environments differ due to causal interventions. Using this causal lens, we propose new machine learning algorithms that achieve robustness to environments unseen in training data, quickly adapt to new envir onments, are robust to measurement drift (as commonly arises in non-stationary data), and use shared structure to improve sample-eff iciency. Instead of assuming that we know the causal structure a priori, we study how one can discover the relevant causal structure as part of learning, also emphasizing that it often suffices to only learnpartial structure. We also study the effect of hidden con founding variables, proposing algorithms that would succeed even in their presence. (Approved for Public Release)

Document Details

Document Type
DoD Grant Award
Publication Date
Aug 20, 2021
Source ID
N000142112807

Entities

People

  • David A Sontag

Organizations

  • Massachusetts Institute of Technology
  • Office of Naval Research
  • United States Navy

Tags

Fields of Study

  • Computer science

Readers

  • Neural Network Machine Learning.

Technology Areas

  • AI & ML
  • AI & ML - Bayesian Inference
  • AI & ML - Neural Networks