Geometry of information flow and uncertainty quantification for robust neural network architectures in deep learning
Abstract
Reliable implementation of machine and deep learning techniques on neural networks requires a fundamentalunderstanding of how and why these algorithms work, and it necessitates determining application-adapted networkarchitecture to give robust predictions/estimates. Innovative network architectures and training algorithms for specificapplications in artificial intelligence, autonomous adaptation, and learning continue to be developed at a fast pace. However,there is no rigorous framework which allows for a systematic understanding of the many successes of machinelearning on neural networks. Crucially, important issues concerning robustness and reliability of such approaches for prediction/estimation from deep neural networks trained on uncertain, sparse data sets remain largely unexplored. Moreover,there exist a number of well-known pitfalls of neural network-based classification and their sensitivity to input perturbations(see Figure ??). Probability theory and information theory furnish the analysis of neural networks with powerfulmethods which allow to uncover the hidden geometry of information flow within a given network architecture. In particular,an approach building on information geometry, data assimilation, and asymptotic statistical theory of model selectionprovide a systematic, and synergistic approach to studying this problem.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 17, 2020
- Source ID
- N629092012037
Entities
People
- Michal Branicki
Organizations
- Office of Naval Research
- United States Navy
- University of Edinburgh