CISE Seminar: Vikas Dhiman, University of Maine

About this event

Vikas Dhiman

Assistant Professor, Electrical and Computer Engineering

University of Maine

An Uncertainty-Aware Safe Controller Is Only as Safe as Its Uncertainty Estimate

Machine learning is increasingly being used in safety-critical scenarios. It is natural to ask what guarantees of safety such systems have and under what assumptions. Control barrier functions (CBFs) are valuable tools for guaranteeing safety even under uncertainty, but only if the uncertainty estimate can be trusted. Theoretically sound uncertainty estimation in neural networks remains an unsolved research problem. In this talk, I will present our work that investigates uncertainty estimation from the perspective of worst-case error bounds in spline neural networks. We find that many of the existing techniques do not use common robotics assumptions such as Lipschitz continuity on the true function for uncertainty estimation. Furthermore, many existing uncertainty estimation techniques are less interpretable because they do not follow a property called distance-awareness. This property demands that the uncertainty estimate from a neural network model at a given test point increase as the test point grows farther in distance from the training data by some distance measure. We address these limitations by building uncertainty estimates from the neuron level to the network level in spline neural networks. Finally, we apply the new uncertainty estimates to CBF-based safe control to make robotic systems safe. While our work estimates uncertainty on spline neural networks, it can be extended to commonly used CNNs and Transformers in the future.

Vikas Dhiman is an Assistant Professor of Electrical and Computer Engineering at the University of Maine, where he directs the Computer Vision and Autonomous Robotics (CVAR) Lab. His research aims to build robots that can operate in unstructured, safety-critical environments with formal guarantees, combining uncertainty quantification in neural networks, learning-enabled safe control, and robot perception. He received his Ph.D. from the University of Michigan in 2019 under Jason Corso, working on localization, mapping, and reinforcement learning for navigation. He was then a postdoctoral researcher with Nikolay Atanasov and Henrik Christensen at UC San Diego. He holds an M.S. from the University at Buffalo and a B.Tech. from IIT Roorkee. His work has appeared in CVPR, IROS, ICRA, AAAI, ICASSP, IEEE TAC, and IEEE TPAMI.

Faculty Host: Roberto Tron

Student Host: Lizzy Yue