Deep Uncertainty

Abstract: Among all the methods intended to estimate the uncertainty of deep network predictions, MC-Dropout and Deep Ensembles are the most widely used. The latter tend to deliver better estimates but at the cost of substantial computational and memory overheads, which makes them unsuitable for many real-world applications.

In this talk, I will discuss alternative approaches that are designed to deliver the performance of Ensembles at a reduced computational cost by taking advantage of the regularities that can be found in the training data. The methods we are proposing take advantage of what if sometimes considered a weakness of neural nets: They behave very differently for in- and out-of-distribution samples, which makes it possible to tell the ones from the others. Furthermore, when physics-based knowledge is available, the network can easily be engineered to exploit it and to deliver improved performance both in terms of accuracy and uncertainty estimation.

Biography Pascal Fua received an engineering degree from Ecole Polytechnique, Paris, in 1984 and the Ph.D. degree in Computer Science from the University of Orsay in 1989. He then worked at SRI International and INRIA Sophia-Antipolis as a Computer Scientist. He joined EPFL in 1996 where he is now a Professor in the School of Computer and Communication Science and heads the Computer Vision Laboratory.His research interests include shape modeling and motion recovery from images, analysis of microscopy images, and machine learning. He has (co)authored over 300 publications in refereed journals and conferences. He is an IEEE Fellow and has been an Associate Editor of IEEE journal Transactions for Pattern Analysis and Machine Intelligence. He often serves as program committee member, area chair, and program chair of major vision conferences and has cofounded three spinoff companies (Pix4D, PlayfulVision, and NeuralConcept).