Uncertainty in Deep Learning: A Probabilistic Robotics Perspective

Abstract

Robots are physical systems that perceive, plan, and act in the real world. As a consequence, their mistakes can not only cause failures in the robots' mission, but they can even endanger human lives, in the case of a robotic surgeon or a self-driving car, for example. This motivates probabilistic robotics, i.e., a paradigm of robotics with a set of methods that enable the robots to assess the uncertainty in their sensory data, used algorithms, learned predictors, etc., such that the robots can plan safe actions. One of the challenges herein is uncertainty quantification in the systems that rely on neural networks. For this, Bayesian statistics provide theoretical foundations. However, bringing Bayesian statistics to neural networks -- referred to as Bayesian Deep Learning -- involves the problem of (a) the choice of well-specified priors, (b) the inference of the posteriors, and (c) the uncertainty estimation through marginalization, which are active areas of research in machine learning and beyond.

For all these sub-problems of Bayesian Deep Learning, this work provides novel methodologies that are well-suited for their applications to robotics. Concretely, we advance the generalization of learning algorithms through priors, the scalability of the inference algorithms to obtain complex posteriors, and the run-time efficiency of marginalization for predictions. We achieve this while improving the quality of uncertainty estimates when compared to existing methods. Inspirations have been drawn from the theories of generalization, information, as well as the universal approximation theorem of neural networks. Theoretical foundations are further provided for all the devised methodologies. With these results, we finally develop probabilistic robotic systems that can improve the performance of deep learning in real-world applications. Starting from (a) a humanoid robot recognizing deformable objects, (b) in-flight aerodynamic analysis of stratospheric and manned helicopter flights, (c) semi-autonomous aerial manipulation at night, to (d) shared autonomy with haptic and extended reality, we provide several system-level contributions.

As a result, the contribution of this thesis stands on two pillars of probabilistic robotics -- one on methodological advances to quantify uncertainty, and another on systems contributions that show possibilities for new applications. Through this lens of probabilistic robotics, we provide a new perspective on the general problem of uncertainty in deep learning.