[Paper Review] Out of Distribution Generalization in Machine Learning

Martín Arjovsky|arXiv (Cornell University)|Mar 3, 2021

Machine Learning and Algorithms68 references50 citations

TL;DR

A comprehensive survey and framework for understanding out-of-distribution generalization in ML, focusing on robustness, invariance, and causal perspectives, with IRM and domain-specific discussions.

ABSTRACT

Machine learning has achieved tremendous success in a variety of domains in recent years. However, a lot of these success stories have been in places where the training and the testing distributions are extremely similar to each other. In everyday situations when models are tested in slightly different data than they were trained on, ML algorithms can fail spectacularly. This research attempts to formally define this problem, what sets of assumptions are reasonable to make in our data and what kind of guarantees we hope to obtain from them. Then, we focus on a certain class of out of distribution problems, their assumptions, and introduce simple algorithms that follow from these assumptions that are able to provide more reliable generalization. A central topic in the thesis is the strong link between discovering the causal structure of the data, finding features that are reliable (when using them to predict) regardless of their context, and out of distribution generalization.

Motivation & Objective

• Clarify what out-of-distribution generalization (OOD) means across problems and applications.
• Present a framework to quantify and compare OOD performance under multiple environments.
• Survey existing methods (robust optimization, domain adaptation, invariance) and analyze their underlying assumptions.
• Highlight the role of causal and invariant features in achieving generalization and identify limitations of current approaches.

Proposed method

• Introduce a framework where OOD risk is the worst-case risk across a set of environments: ROOD(f)=max_e R^e(f).
• Discuss training environments and the need for E_train to be related to the broader environment set E_all.
• Provide motivating examples illustrating spurious correlations and how they affect OOD performance.
• Examine connections between empirical risk minimization (ERM), robust optimization, and Wasserstein/KL robustness.
• Develop and reference Invariant Risk Minimization (IRM) and its nonlinear variants as approaches to enforce invariances across environments.
• Outline domain-specific case studies and practical implications across robotics, vision, NLP, and fairness.

Experimental results

Research questions

• RQ1 How can we quantify and compare out-of-distribution generalization across different problems?
• RQ2 Under what assumptions do various OOD methods (robust optimization, domain adaptation, invariance) succeed or fail?
• RQ3 What is the role of invariant or causal features in achieving robust generalization across environments?
• RQ4 How can we relate training environments to unseen test environments to achieve better OOD performance?
• RQ5 What are the limitations and future directions for OOD generalization in real-world applications?

Key findings

• There is no single general OOD method; performance depends on problem-specific assumptions about environments.
• A framework based on worst-case risk over environments provides a principled way to measure OOD performance.
• Invariance-based methods, such as IRM, aim to identify predictors that are stable across environments, aiding generalization when such invariances exist.
• Empirical discussions include synthetic and colored MNIST experiments illustrating spurious correlations and robustness challenges.
• Domain-specific perspectives show how OOD generalization considerations differ across robotics, computer vision, NLP, and fairness settings.
• The thesis emphasizes the trade-offs between realism of assumptions and practical gains in OOD generalization.