[Paper Review] OoD-Bench: Benchmarking and Understanding Out-of-Distribution Generalization Datasets and Algorithms

Nanyang Ye, Kaican Li+5 more|arXiv (Cornell University)|Jun 7, 2021

Domain Adaptation and Few-Shot LearningComputer Science85 references22 citations

TL;DR

This paper introduces OoD-Bench, a unified benchmark for evaluating out-of-distribution (OoD) generalization in deep learning. It identifies two pervasive distribution shifts, evaluates existing OoD algorithms across these shifts, and reveals that algorithms strong on one shift often underperform on the other, highlighting the need for more robust, generalizable methods.

ABSTRACT

Deep learning has achieved tremendous success with independent and identically distributed (i.i.d.) data. However, the performance of neural networks often degenerates drastically when encountering out-of-distribution (OoD) data, i.e., training and test data are sampled from different distributions. While a plethora of algorithms has been proposed to deal with OoD generalization, our understanding of the data used to train and evaluate these algorithms remains stagnant. In this work, we position existing datasets and algorithms from various research areas (e.g., domain generalization, stable learning, invariant risk minimization) seemingly unconnected into the same coherent picture. First, we identify and measure two distinct kinds of distribution shifts that are ubiquitous in various datasets. Next, we compare various OoD generalization algorithms with a new benchmark dominated by the two distribution shifts. Through extensive experiments, we show that existing OoD algorithms that outperform empirical risk minimization on one distribution shift usually have limitations on the other distribution shift. The new benchmark may serve as a strong foothold that can be resorted to by future OoD generalization research.

Motivation & Objective

• To unify and systematize disparate research on OoD generalization from domains like domain generalization and invariant risk minimization.
• To identify and quantify two distinct types of distribution shifts prevalent across existing datasets.
• To evaluate the performance of existing OoD generalization algorithms under a new, comprehensive benchmark dominated by these two distribution shifts.
• To expose the limitations of current algorithms that perform well on one type of distribution shift but fail on the other.

Proposed method

• The authors analyze existing OoD generalization datasets to identify and characterize two dominant types of distribution shifts.
• They design a new benchmark, OoD-Bench, that systematically incorporates these two distribution shifts across diverse data distributions.
• They evaluate a wide range of state-of-the-art OoD generalization algorithms on this benchmark using standardized protocols.
• They compare algorithm performance across the two distribution shifts to reveal trade-offs and limitations in generalization capacity.

Experimental results

Research questions

• RQ1 What are the two dominant types of distribution shifts that underlie existing OoD generalization datasets and algorithms?
• RQ2 How do existing OoD generalization algorithms perform when evaluated under a unified benchmark that emphasizes these two distribution shifts?
• RQ3 Are algorithms that outperform empirical risk minimization on one type of distribution shift equally effective on the other?
• RQ4 To what extent do current OoD algorithms generalize across different types of distribution shifts?

Key findings

• The two identified distribution shifts—domain shift and concept shift—are pervasive across existing OoD datasets and fundamentally shape model generalization behavior.
• Existing OoD algorithms that outperform empirical risk minimization on one distribution shift often fail to generalize to the other, indicating a critical gap in robustness.
• No single algorithm consistently outperforms empirical risk minimization across both types of distribution shifts, revealing a fundamental limitation in current approaches.
• The proposed OoD-Bench benchmark effectively exposes these performance trade-offs and serves as a strong foundation for future OoD generalization research.