Traditional machine learning models are based on the assumption that both training and test data are Independent and Identically Distributed (i.i.d.). However, in real-world applications, this i.i.d. assumption often fails to hold due to unforeseen distributional shifts, leading to considerable degradation in model performance even if in-domain accuracy is excellent. This project focuses on the complex setting called Out-Of-Distribution Generalization (OOD generalization), where the distributions of the test data differ from those of the training data. The aim of the project is to build a coherent Learning theory for learning under domain shift, with a specific focus on studying the domain-shift complexity term inside generalization bounds. More specifically, I will develop a data-driven notion of domain-shift complexity that can be estimated in domain adaptation settings, extend it to domain generalization, and investigate how simple causal or invariance assumptions can be used to control this quantity. These theoretical developments will be complemented by implementing the corresponding learning objectives or regularisers and empirically evaluating them on standard benchmark datasets.
Statistical learning theory, Mathematically explainable machine learning, Causal inference
BSc in Financial Mathematics, Tianjin University
MSc in Financial Technology with Data Science, University of Bristol
Dr Mingzhi Dong
Prof Michael Yang
