Group Robust Classification Without Any Group Information

Christos Tsirigotis, Joao Monteiro, Pau Rodriguez, David Vazquez, Aaron Courville

Abstract

Standard machine learning models often rely on spurious correlations, leading to poor performance on minority groups or out-of-distribution data. While group robustness techniques aim to mitigate this, they typically require expensive human-annotated bias labels for training or model selection. This paper introduces Unsupervised Logit Adjustment (uLA), a two-phase framework that achieves group robustness without any manual group information. By leveraging self-supervised pretraining to identify biases and employing a novel unsupervised validation criterion, uLA matches or exceeds the performance of state-of-the-art methods that rely on supervised labels.

Introduction

Empirical Risk Minimization (ERM), the standard training paradigm for most neural networks, is fundamentally "lazy." It tends to prioritize the easiest path to minimizing loss, which often involves latching onto spurious correlations—shortcuts that hold true in the majority of training data but fail in the real world. A classic example is a model that learns to identify a "camel" based on the presence of "sand" rather than the animal's physical features. When presented with a camel in a forest, the model fails.

To address this, researchers focus on group robustness, ensuring models perform well across all subgroups (e.g., camels on sand and camels on grass). However, existing solutions face two significant hurdles:

  1. The Annotation Bottleneck: Most robust methods require "bias labels" (e.g., identifying the background of every image) to train the model or to perform hyperparameter tuning during validation.
  2. The Systematicity Gap: Many methods assume that every combination of attributes (e.g., every animal on every background) appears at least once in the training set. In reality, many combinations are entirely missing, a challenge known as systematic generalization.

The uLA Framework

The proposed Unsupervised Logit Adjustment (uLA) framework addresses these challenges through a two-phase process that requires no group labels at any stage.

Phase 1: Extracting a Biased Proxy

The first step involves identifying the biases inherent in the dataset. The authors utilize Self-Supervised Learning (SSL) (such as MoCoV2+) to extract rich feature representations without using labels. They then train a low-capacity linear classifier on top of these frozen features to predict target labels. Because the linear model is intentionally limited in capacity, it naturally gravitates toward the most prominent spurious correlations, effectively becoming a "biased proxy" that mirrors the dataset's flaws.

Phase 2: Debiasing via Logit Adjustment

Using the predictions from the biased proxy, the framework estimates a "soft confusion matrix" to identify which target-bias combinations are overrepresented. The main model is then finetuned using Logit Adjustment. This technique additively shifts the model's output scores (logits) during training to suppress the influence of the spurious correlations identified by the proxy, forcing the model to learn more intrinsic features.

Unsupervised Validation

A critical contribution of uLA is its approach to model selection. Many "unsupervised" methods in existing literature are actually semi-supervised because they use group labels to pick the best model during validation. uLA breaks this dependency by using the biased proxy to define a Balanced Accuracy validation criterion, allowing for truly unsupervised hyperparameter tuning.

Experimental Evaluation

The authors evaluated uLA across synthetic and real-world benchmarks:

  • Synthetic: CMNIST and CCIFAR10, where objects are correlated with specific colors or textures.
  • Real-World: Waterbirds (birds vs. backgrounds) and CelebA (gender vs. hair color).
  • Systematic Generalization: The authors introduced SMPI3D (Systematic MPI3D), a benchmark using robotic arm images to test unseen attribute combinations (e.g., identifying a "blue cube" when only "red cubes" and "blue spheres" were seen during training).

Key Results and Insights

Performance and Robustness

uLA demonstrates that bias labels are not a prerequisite for robustness. On the Waterbirds dataset, uLA achieved a worst-group accuracy of 86.1%, approaching the 91.4% achieved by GroupDRO, which requires full bias labels during training. On synthetic datasets like CMNIST, uLA significantly outperformed ERM and rivaled state-of-the-art methods that utilize bias labels for validation.

Success in Systematic Generalization

The SMPI3D benchmark revealed a significant weakness in current robust algorithms: most perform worse than basic ERM when faced with entirely missing attribute combinations. uLA was the only method to consistently improve performance in this extreme scenario. On the most difficult version of the task, uLA achieved 59.58% accuracy, while ERM and other methods struggled at approximately 31%.

The Role of SSL

A major insight of the paper is that SSL-pretrained features are exceptionally good at identifying biases. A simple linear probe on these features acts as a "canary in the coal mine," highlighting exactly where the dataset is skewed and providing the necessary information for debiasing without human intervention.

Conclusion and Future Work

uLA provides a path forward for training robust models in domains where bias labels are expensive, unavailable, or ethically problematic to collect. While a performance gap remains between uLA and fully supervised methods on complex tasks, the framework proves that unsupervised bias extraction is a viable and powerful strategy. Future research may focus on more sophisticated methods for extracting bias from SSL representations and applying these techniques to high-stakes fairness applications.

Summary updated

Abstract

Empirical risk minimization (ERM) is sensitive to spurious correlations present in training data, which poses a significant risk when deploying systems trained under this paradigm in high-stake applications. While the existing literature focuses on maximizing group-balanced or worst-group accuracy, estimating these quantities is hindered by costly bias annotations. This study contends that current bias-unsupervised approaches to group robustness continue to rely on group information to achieve optimal performance. Firstly, these methods implicitly assume that all group combinations are represented during training. To illustrate this, we introduce a systematic generalization task on the MPI3D dataset and discover that current algorithms fail to improve the ERM baseline when combinations of observed attribute values are missing. Secondly, bias labels are still crucial for effective model selection, restricting the practicality of these methods in real-world scenarios. To address these limitations, we propose a revised methodology for training and validating debiased models in an entirely bias-unsupervised manner. We achieve this by employing pretrained self-supervised models to reliably extract bias information, which enables the integration of a logit adjustment training loss with our validation criterion. Our empirical analysis on synthetic and real-world tasks provides evidence that our approach overcomes the identified challenges and consistently enhances robust accuracy, attaining performance which is competitive with or outperforms that of state-of-the-art methods, which, conversely, rely on bias labels for validation.
BibTeX
@inproceedings{tsirigotis2023group,
  author = {Christos Tsirigotis and Joao Monteiro and Pau Rodriguez and David Vazquez and Aaron Courville},
  title = {{Group Robust Classification Without Any Group Information}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2023},
  url = {https://openreview.net/forum?id=2OcNWFHFpk},
}
Figure