Distribution shift in machine learning: robustness, invariance, and a causal view – Wooseok Ha
December 29 @ 4:30 pm - 6:30 pm KST
https://www.ibs.re.kr,
55 Expo-ro Yuseong-gu
Daejeon, Daejeon 34126 Korea, Republic of + Google Map
Daejeon, Daejeon 34126 Korea, Republic of + Google Map
Classical machine learning models are typically trained under the assumption that the training (source) and test (target) data are drawn from the same distribution. However, real-world data are rarely clean or consistent, and distribution shifts between the source and target domains are ubiquitous. Despite its importance, addressing distribution shifts is highly difficult. The fundamental challenge is that the problem is mathematically ill-posed: shifts can occur in many different forms, and no single method can handle all of them. While numerous algorithms have been proposed in recent years to solve distribution shifts, most are empirical-driven and lack solid foundations. In this talk, I will provide a broad overview of approaches to address distribution shift based on invariance and distributional robustness, and explain how these methods are intrinsically connected to a causal perspective. In particular, I will show why it is crucial to carefully formulate assumptions that relate the source and target domains for reliable generalization, and how assumptions grounded in the causal system enable the analysis of algorithms under both unsupervised and semi-supervised settings.