Uncertainty Quantification of ML models: Hands-on Introduction

[psteinb]

Uncertainty Quantification of ML models: From Introduction to Advanced

Responsible person(s)

Sebastian Starke, , HZDR,
Steve Schmerler, HZDR, @elcorto
Peter Steinbach, HZDR, @psteinb

Gianni Franchi, ENSTA Paris, @giannifranchi
Olivier Laurent ENSTA Paris and Paris Saclay University, @OLaurent

Format

Tutorial and Workshop

Timeframe

• 13:30h Introduction to Uncertainties by Peter Steinbach
• break
• 15:00h Gianni Franchi et al:
  • packed ensembles to extend ensembles (15min intro + 15min hands on) -> 45'
  • Bayesian Neural Networks with torch_uncertainty (45min intro + 45min hands on) -> 120'
• 17:30-18:00h Finish

Description

In this tutorial, we will give a hands-on introduction to uncertainty quantification for ML models. We will focus on MCDropout and DeepEnsembles as the traditional methods used in the field in the beginning. We will then turn to more advanced topics like Bayesian Neural Networks and accelerated Deep Ensembles. We are super happy to received support by the torch_uncertainty team. The workshop itself will offer a mixture of teaching presentations and exploratory exercises using local or remote notebooks. We are planning enough time for all participants to ask questions.

Requirements

Each beginner is expected to bring their laptop with a working python interpreter (at best python 3.10 or 3.11).

[SusanneWenzel]

Thanks for this contribution @psteinb! Fits the overall programme very good! I'm happy to help setting it up. Do you have an (initial) idea about the number of people you can handle?
Do you plan an open call to get the speakers or will you invite specific experts?

[psteinb]

So, the number of people depends on how well we advertise the workshop. ;-) Jokes aside, we will be 5 people supporting the workshop. So anything up to 40 people is doable from my point of view. But my honest estimate would be 20 participants.

[psteinb]

We have already a tentative agenda and 2 external speakers. Should I drop everything here?

[SusanneWenzel]

@psteinb yes, please update the agenda here.