AILo Talk - Ivo de Jong, University of Groningen

Title: Introducing myself and Uncertainty Quantification

Abstract:

Machine Learning suffers from overconfidence. Simple ML models make a prediction without any indication of how likely this prediction is to be correct. As a result it can look like an ML model makes "dumb" mistakes, but if you give it the opportunity to say "I don't know", you may be able to avoid these misclassifications. Uncertainty Quantification gives a measure of uncertainty with each prediction from a Machine Learning model, so you know when the prediction can be trusted and when it cannot. Additionally, some Uncertainty Quantification methods can even do this on "unfamiliar" data, allowing it to detect when testing samples are different from the training data. In this AILo talk I'll present intuitively how these uncertainty quantification methods work, and what they may be used for, in an attempt to plant notions of uncertainty in your brain. I'm aiming for an informal presentation, so prepare to interrupt me with comments about puppies, estimations of heteroscedastic uncertainty, and the weather.