Towards Out-Of-Distribution Generalization: A Survey (2108.13624v2)

Abstract: Traditional machine learning paradigms are based on the assumption that both training and test data follow the same statistical pattern, which is mathematically referred to as 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 upon deployment. This observed discrepancy indicates the significance of investigating the Out-of-Distribution (OOD) generalization problem. OOD generalization is an emerging topic of machine learning research that focuses on complex scenarios wherein the distributions of the test data differ from those of the training data. This paper represents the first comprehensive, systematic review of OOD generalization, encompassing a spectrum of aspects from problem definition, methodological development, and evaluation procedures, to the implications and future directions of the field. Our discussion begins with a precise, formal characterization of the OOD generalization problem. Following that, we categorize existing methodologies into three segments: unsupervised representation learning, supervised model learning, and optimization, according to their positions within the overarching learning process. We provide an in-depth discussion on representative methodologies for each category, further elucidating the theoretical links between them. Subsequently, we outline the prevailing benchmark datasets employed in OOD generalization studies. To conclude, we overview the existing body of work in this domain and suggest potential avenues for future research on OOD generalization. A summary of the OOD generalization methodologies surveyed in this paper can be accessed at http://out-of-distribution-generalization.com.

Towards Out-Of-Distribution Generalization: A Survey

The paper "Towards Out-Of-Distribution Generalization: A Survey" by Jiashuo Liu et al. represents a comprehensive examination of the Out-of-Distribution (OOD) generalization problem, a significant challenge in machine learning where models suffer performance degradation due to distributional shifts between training and test data. The authors aim to systematically survey existing methodologies, propose a coherent categorization, and provide insights into the future trajectory of OOD generalization research.

The paper begins by defining the OOD generalization problem within the context of traditional machine learning's $i.i.d.$ assumption. The authors assert that real-world applications often exhibit distributional shifts rendering conventional supervised learning approaches inadequate. Addressing this gap, they delineate a structured survey encompassing unsupervised representation learning, supervised model learning, and optimization methodologies, each contributing uniquely to the broader OOD generalization challenge.

For unsupervised representation learning, the focus is on techniques like unsupervised domain generalization and disentangled representation learning. These aim to develop representations independent of domain-specific variations. Disentangled representation learning, particularly causal representations, shows potential by separating latent factors of variation, although the complexity of these methodologies warrants further exploration.

Supervised model learning methodologies are extensively discussed, emphasizing invariant learning, causal learning, and innovative training strategies. Invariant Risk Minimization (IRM) and Domain-Adversarial Training strategies underscore attempts to cultivate generalization by ensuring model invariance across different domains. The paper highlights both achievements and limitations in these domains, suggesting that while invariant learning methods hold promise, they often struggle with practical limitations such as the requirement for multiple training environments or the pitfalls revealed in empirical assessments.

Optimization for OOD involves leveraging distributionally robust optimization (DRO) techniques, which aim for robust performance against worst-case distributional shifts. These methods include formulating uncertainty sets using divergence measures or Wasserstein distances. The survey critically evaluates the efficacy of these methodologies, recognizing the ongoing challenge of balancing robustness with practical feasibility, as overly conservative approaches may hinder model accuracy in real-world scenarios.

For empirical evaluation, the paper underscores the need for appropriate datasets and benchmarks to assess OOD generalization capabilities reliably. The authors highlight pertinent challenges, such as the inadequacy of existing benchmarks and the necessity for studies that elucidate the conditions under which current approaches succeed or fail.

The paper also touches on the overlapping concerns of fairness and explainability within the context of OOD generalization. Fairness is intrinsically related to handling subgroups akin to environments, with DRO strategies historically linked to fairness guarantees under distribution shifts. Explainability aligns with causal inference approaches, suggesting that OOD generalization and model interpretability might be pursued concurrently through causally informed methodologies.

In conclusion, the survey identifies critical areas for future exploration in OOD generalization, such as better theoretical characterizations, practical environment demands, and more suitable evaluation protocols. The incorporation of pre-trained LLMs introduces a novel dimension to this landscape, with the need for understanding their OOD capabilities becoming increasingly significant as these models dominate various AI applications. This paper, thus, not only serves as a cornerstone reference for researchers exploring OOD generalization but also delineates a path for future investigations in this evolving field.