We will show how to assess the uncertainty of deep neural networks. We will cover Bayesian Deep Learning and other out-of-distribution detection methods. The talk will include examples that will show how to implement the methods in Pytorch.
Overconfidence is nowadays one of the most hot topics in discussion about safety of Machine Learning applications. This problem strikes both, models which very often suffer from providing confident scores for problematic cases, and researchers who are often likely to believe that good results on validation sets prove that model will generalize well to the new examples. Because of this we would like to dedicate our talk to the following topics:
Analysis of different kinds of problems that may occur due to overconfidence and their corresponding uncertainty types,
Introduction of a variety of methods on how to estimate the level of uncertainty for a given model. We will mostly concentrate on techniques from bayesian deep learning family.
We will show how to implement them with Pytorch.