Bridging Distribution Shift and AI Safety: Conceptual and Methodological Synergies

TL;DR

This paper bridges distribution shift and AI safety through a comprehensive analysis of their conceptual and methodological synergies, establishing two types connections between specific causes of distribution shift and fine-grained AI safety issues.

Abstract

This paper bridges distribution shift and AI safety through a comprehensive analysis of their conceptual and methodological synergies. While prior discussions often focus on narrow cases or informal analogies, we establish two types connections between specific causes of distribution shift and fine-grained AI safety issues: (1) methods addressing a specific shift type can help achieve corresponding safety goals, or (2) certain shifts and safety issues can be formally reduced to each other, enabling mutual adaptation of their methods. Our findings provide a unified perspective that encourages fundamental integration between distribution shift and AI safety research.

Figures (11)

• Figure 1: Connections between distribution shift causes and AI safety issues. Dashed arrows: Methods addressing a specific cause of distribution shift can also achieve a particular safety goal. Bidirectional solid arrows: Due to inherent consistency between a certain distribution shift cause and a corresponding safety issue, the two can be reduced from one another and methods derived for them can be mutually adapted.
• Figure 2: Summary of safety issues related to distribution shift covered in this paper.
• Figure 3: Illustration of the inclusion relationships among different causes of distribution shift.
• Figure 4: An illustration of individual selection bias and group selection bias. The blue curves represent the input distribution of the source domain, and the green curves represent the input distribution of the target domain. For individual selection bias, the two distributions are known to differ. However, it is hard to attribute this difference to a certain subpopulation in the data. In contrast, for group selection bias, apparent changes in different subpopulations causes the distribution shift.
• Figure 5: Data pruning is a dual problem of addressing individual selection bias. To address individual selection bias, methods focus on refining the training process so that a model trained on source data performs well on target data. As a comparison, data pruning methods prunes less informative data from the original dataset, so that a model trained on the remaining subset still performs well. Both problems share the same goal of improving test accuracy under distribution differences.

Theorems & Definitions (17)

• Definition 1: Individual Selection Bias
• Definition 2: Data Pruning
• Definition 3: Group Selection Bias
• Definition 4: Risk Parity
• Definition 5: Environmental Change
• Definition 6: Sufficiency, Independence, and Separation
• Definition 7: Test Fairness
• Definition 8: Well-calibration
• Definition 9: Demographic Parity
• Definition 10: Equalized Odds