New information metrics and stochastic optimization methods for robust and trustworthy statistical learning

Abstract

Artificial intelligence models are often intrinsically random and therefore described by probability distributions. The PIs propose to build new information-theoretic objects (called divergences) that quantify dissimilarity between models or data sets. The versatility and mathematical properties of these novel divergences will allow to build new, adaptive and physics-informed learning algorithms. These new game-theoretic learning algorithms (called generative adversarial networks) which combine neural networks with the new divergences will be a powerful tool to construct artificial intelligence models from data and expert knowledge. As a demonstration we construct new data-infused multi-scale, multi-physics models for the design of more efficient fuel cells. These new divergences can also be used as an uncertainty quantification tool for complex, artificial intelligence models. Stress tests can be constructed to probe and quantify which component of the model is sensitive to uncertainties in the nature of the model itself and which component will benefit from added investment into more reliable data. This allows for systematic and efficient design of more robust and trustworthy artificial intelligence models.

Document Details

Document Type
DoD Grant Award
Publication Date
Mar 07, 2023
Source ID
FA95502110354

Entities

People

  • Paul Dupuis

Organizations

  • Air Force Office of Scientific Research
  • Brown University
  • United States Air Force

Tags

Fields of Study

  • Computer science

Readers

  • Distributed Systems and Data Platform Development
  • Neural Network Machine Learning.
  • Theoretical Analysis.

Technology Areas

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