An Information-theoretic Approach to Distribution Shifts

Marco Federici, Ryota Tomioka, Patrick Forré

Introduction

Standard machine learning paradigms typically rely on the Independent and Identically Distributed (IID) assumption, where training and testing data are drawn from the same underlying distribution. However, real-world applications frequently encounter distribution shifts—scenarios where a model trained in one environment (e.g., a specific hospital or geographic region) fails to generalize to another due to changes in equipment, demographics, or environmental conditions.

The primary obstacle in achieving robustness is the model's tendency to exploit "spurious correlations"—features that are predictive within the training set but do not maintain a stable relationship with the target variable across different domains. While various invariance criteria have been proposed to mitigate this, a unified understanding of their relative strengths and applicability has been lacking. This paper introduces an information-theoretic framework to decompose prediction error and evaluate regularization strategies across diverse causal structures.

Conceptual framework of distribution shift and the underlying causal selection mechanism.

Information-Theoretic Decomposition

The authors propose a framework based on Mutual Information to analyze the components of error under shift. By introducing a selection variable $t$ (where $t=1$ denotes training data and $t=0$ denotes test data) and a latent representation $z$ of the input features $x$, the total prediction error can be decomposed into two fundamental parts:

  1. Information Loss: The error incurred when the encoder discards predictive information from $x$ that would have been useful for determining $y$.
  2. Latent Test Error: The error resulting from "concept shift," where the relationship between the representation $z$ and the target $y$ changes between the training and test environments.

This decomposition allows for a rigorous analysis of common regularization families, including the Information Bottleneck, Independence (Statistical Parity), Sufficiency (Predictive Parity), and Separation (Equalized Odds).

Causal Structures and Experimental Setup

To test the hypothesis that the success of invariance criteria depends on the underlying data-generating process, the authors utilized variants of the Colored MNIST (CMNIST) dataset. In the standard CMNIST, color serves as a spurious predictor that correlates with the label during training but flips during testing.

Two additional variants were introduced to simulate more complex causal dependencies:

  • d-CMNIST: Introduces a dependency between the environment $e$ and the digit $d$ itself.
  • y-CMNIST: Introduces a dependency between the environment $e$ and the label $y$.

Empirical Results and Insights

The empirical analysis reveals several key findings:

  • No Universal Invariance: No single regularization objective is optimal across all scenarios. For example, the "Separation" criterion performs exceptionally well on standard CMNIST but fails on d-CMNIST because it cannot account for direct dependencies between the environment and features.
  • The Invariance-Information Trade-off: Removing bias requires discarding information. If a model discards too much, "Information Loss" increases, leading to poor overall performance. The challenge lies in identifying the "maximally informative invariant representation."
Causal graphs for CMNIST variants and the resulting decomposition of test error across different regularization objectives.
  • Causal Structure as a Guide: The effectiveness of a model selection procedure is intrinsically linked to the causal structure of the data. Without knowledge of how data was selected or biased, choosing a regularization objective is often arbitrary.
  • Optimization Challenges: Theoretical guarantees for methods like Invariant Risk Minimization (IRM) often fail to materialize in practice due to the difficulty of optimizing the resulting loss functions and their sensitivity to hyperparameters.

Conclusion

This work provides a unifying framework for understanding distribution shifts through the lens of information theory. By demonstrating that the utility of invariance criteria is context-dependent, the authors highlight the necessity of considering the underlying causal graph when designing robust machine learning systems. Future work may leverage this error decomposition to assist in causal discovery, using patterns of model failure to infer the structure of the environment.

Summary updated

BibTeX
@inproceedings{federici2021information,
  author = {Marco Federici and Ryota Tomioka and Patrick Forré},
  title = {{An Information-theoretic Approach to Distribution Shifts}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2021},
  url = {https://openreview.net/forum?id=GrZmKDYCp6H},
}
Figure