Title: A Causal Approach to Transfer Learning in Machine Learning

Abstract:
A fundamental challenge in AI is ensuring performance guarantees for predictions in unseen domains. In reality, there can be significant uncertainty about the distribution of new data and how well existing predictors will perform. For example, a risk prediction tool fine-tuned for a specific patient population (e.g., a particular hospital or geographic region) may not perform optimally when deployed to a different patient population with varying characteristics. This talk explores the problem through the lens of partial transportability, which combines data from different sources. By encoding assumptions about data-generating mechanisms in causal diagrams, we can provide guarantees for a classifier’s out-of-distribution performance. We can consistently predict the worst-case performance of existing classifiers and, under our assumptions, explicitly train classifiers to optimize for the worst-case performance in the target domain. Both methods can be parameterized with expressive neural networks and implemented using gradient-based optimization. These findings offer a new perspective on transfer learning and domain generalization in machine learning.

Bio:
Alexis Bellot is a Research Scientist at Google DeepMind in London, UK. He previously worked as a postdoctoral researcher with Professor Elias Bareinboim at Columbia University. Before joining Columbia, he earned his Ph.D. in Applied Mathematics from the University of Cambridge under the supervision of Professor Mihaela van der Schaar. Alexis's research focuses on causality in data and its applications, particularly combining causality and machine learning to ensure the safety and alignment of AI systems.