Performance guarantees go beyond empirical accuracy and aim to answer how well a learned model will generalize to new, unseen data. Classical learning theory provides some insights here, but often falls short when applied to deep learning models. Modern approaches seek new ways to explain and guarantee good generalization by analyzing the geometry of the model’s behavior or using probabilistic reasoning tools.

Privacy becomes essential when machine learning is applied to sensitive data, especially in domains like medicine. Simply removing names or IDs is not enough to protect individual information. Instead, we need principled ways to learn from data without exposing it. Key strategies are to ensure that individual contributions to the training data cannot be traced, e.g., by perturbing data before sharing it, or to design training methods that do not require centralizing sensitive data in the first place.

Causality addresses the question of why a model makes a certain prediction, and how robust that prediction is under changes to the data or the environment. Unlike standard prediction-based approaches, causal reasoning aims to uncover underlying relationships and structure in the data, which can help build models that are more interpretable and more stable under distribution shifts. Various techniques allow us to estimate causal structure directly from observational data.

In this Proseminar, participants will learn about methods that tackle these key aspects by surveying a number of key publications and critically comparing them.