BLOOD: Bi-level Learning Framework for Out-of-distribution Generalization

Jun-Hyun Bae, Inchul Choi, Minho Lee

Keywords: Out-of-Distribution Generalization, Generalization, Spurious Correlations, Bi-level Optimization

Abstract: Empirical risk minimization (ERM) based machine learning algorithms have suffered from weak generalization performance on the out-of-distribution (OOD) data when the training data are collected from separate environments with unknown spurious correlations. To address this problem, previous works either exploit prior human knowledge for biases in the dataset or apply the two-stage process, which re-weights spuriously correlated samples after they were identified by the biased classifier. However, most of them fail to remove multiple types of spurious correlations that exist in training data. In this paper, we propose a novel bi-level learning framework for OOD generalization, which can effectively remove multiple unknown types of biases without any prior bias information or separate re-training steps of a model. In our bi-level learning framework, we uncover spurious correlations in the inner-loop with shallow model-based predictions and dynamically re-group the data to leverage the group distributionally robust optimization method in the outer-loop, minimizing the worst-case risk across all batches. Our main idea applies the unknown bias discovering process to the group construction method of the group DRO algorithm in a bi-level optimization setting and provides a unified de-biasing framework that can handle multiple types of biases in data. In empirical evaluations on both synthetic and real-world datasets, our framework shows superior OOD performance compared to all other state-of-the-art OOD methods by a large margin. Furthermore, it successfully removes multiple types of biases in the training data groups that most other OOD models fail.

One-sentence Summary: In this paper, we propose a novel bi-level learning framework for out-of-distribution generalization, which aims to eliminate multiple types of unknown biases.