Control and Learning Enabled Verifiable Robust AI (CLEVR-AI)
Abstract
We expect to carry out the proposed activities according to the following timeline:Year 1 will establish the basic framework for all tasks. A kick-off workshop will bring all participants up to speed on the required background. T1 will establish the basic framework for studying benign overfitting for linear time series and limitations of RNNs as controllers. T2 will start developing a bio inspired unified framework for joint sparse sensing and actuation, based on reverse engineering of strategies observed in our biological testbed. T3 will set the stage to study classes of parametrizations that can be learned tractably, including Koopman based ones. T4 will start out by developing theory to establish uncertainty bounds for given verification goals.Year 2 will extend the basic framework developed in year 1 by considering more complex scenarios and removing some assumptions. T1 will expand the benign overfitting framework to dynamical systems and establish sample complexity results for RNNs. T2 will continue to work on joint sensing and actuation, by developing strategies for systems that are unobservable at a single point but observable under active sensing. T3 will start the development of a theory of identifiability ofNNs. T4 will concentrate on statistical enforcement of performance bounds during training. Year 3 will start bringing into focus practical issues related to using AI in the loop, with the goal of setting the stage for implementations compatible with computational and power budgets available in different scenarios (e.g. single and multiple autonomous vehicles). T1 will extend overfitting to RKHS models and wide RNNs and explore the use of optimal control for training NNs. T2 will focus on sensor activation strategies. T3 will analyze the use of learned constraints for designing more frugal equivalents of a given trained NN and on special purpose compact layers. T4 will concentrate on complexity trade-offs for global verification from local data. Year 4 will focus on the trade-offs between the main factors that influence AI enabled control of dynamical systems with emphasis on the effects of uncertainty. T1 will study the regularization properties of stochastic gradient descent and input to state stability characterization of performance degradation under adversarial attacks. T2 will study the effect of uncertainty and attacks on sensing performance using graph theoretic and information based complexity tools. T3 will analyze the use of learned constraints and statistics based layers for improved robustness. T4 will develop theory and algorithms for density-based robust viable set learning.Year 5 will focus on wrapping up the theoretical framework and connecting results across thrusts. It is expected to lead to an initial draft of a textbook on AI enabled control.Deliverables include technical publications, software (including case studies), the instructional material used in the workshops and short courses, and an initial draft of a textbook covering the fundamentals of AI enabled control of dynamical systems.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 05, 2021
- Source ID
- N000142112431
Entities
People
- Mario Sznaier
Organizations
- Northeastern University
- Office of Naval Research
- United States Navy