Practical Uncertainty in Machine Learning

Date

Nov 22, 2021

Time

1:30 PM - 3:00 PM

Speaker

Prof. Philipp Hennig

Affiliation

University Tübingen / TUEAI

Language

en

Main Topic

Informatik

Host

ScaDS.AI Dresden/Leipzig

Description

Like any good scientist, a decent machine learning method should be able to estimate its own error. Such quantified uncertainty has many uses beyond the basic error bar: It provides the principled mechanisms to guide exploration and active learning, motivate and critique design choices, and trade off the utility of information from multiple sources. Probability Theory provides the universal and rigorous framework to quantify and manipulate uncertainty. The application of this formalism — Bayesian inference — has a reputation to be complicated and expensive. This tutorial will try to dispel this myth. Starting from basic examples we will get to know the Gaussian case a practically-minded workhorse of Bayesian inference, which maps the abstract notions of probability theory onto basic linear algebra. We will then see that modern automatic differentiation tools allow us to transfer this rich language to virtually all of modern machine learning. In particular, we will see how quantified uncertainty can be constructed simply in deep learning, at low computational and implementation overhead.

Links

• Program and registration

Last modified: Nov 24, 2021, 8:26:37 AM