Transfer & Physics Informed Learning

Lectures by Mathilde Mougeot, Antoine de Mathelin and Mounir Atiq

The course on transfer learning (TL) and domain adaptation (DA) addresses the critical challenge of domain shift often encountered in industry, where machine learning models are deployed on data which are not identically distributed according to the training distribution used to build the model. Through a mix of lectures and Python-based tutorials, we explore diverse strategies for adapting the training distribution and the models to a new target distribution.
Lesson 1 presents the importance of transfer learning and domain adaptation through various applications.
Lesson 2 covers importance weighting, a technique that reweights the training distribution to better align with the target distribution.
Lesson 3 explores representation-based approaches that aim to find a shared representation space for both training and target distributions, improving model adaptability.
Lesson 4 discusses model-based transfer, where pretrained models are fine-tuned to perform on new target domains. Tutorials will make use of the Adapt library, allowing participants to implement and experiment with these methods hands-on.
Lesson 5 presents Physics Informed Machine Learning to incorporate knowledge of physical equation in machine learning models.

Conformal Prediction

Lectures by Margaux Zaffran

This course is a beginner-friendly yet detailed introduction to conformal prediction. It will be accompanied by practical coding exercises.
Starting from the basics and most understandable case of split conformal prediction in the standard mean regression setting, passing through an adaptation to quantile regression, we will reach its general version which includes classification.
In a second part, we will present the original approach, namely full conformal prediction, which historically preceded the split version. Intractable in practice, we will also introduce research works based on k-fold strategies, that is in between full and split version. We will discuss its theoretical guarantees insisting not only on their strengths but also on their inherent limits.
In the last part, we will explore how to extend conformal prediction when its only assumption (data exchangeability) is not satisfied. Finally, we will review specific applications of conformal prediction that have been deployed in real life, in order to analyze what advantages conformal brought, but also what limitations it imposed.
At the end of the course, the audience will have the tools to understand when and how to use conformal prediction, but also what are the current active research directions.