Abstract
In subpopulation shift scenarios, a Curriculum Learning (CL) approach would only serve to imprint the model weights, early on, with the easily learnable spurious correlations featured. To the best of our knowledge, none of the current state-of-the-art subpopulation shift approaches employ any kind of curriculum. To overcome this, we design a CL approach aimed at initializing the model weights in an unbiased vantage point in the hypothesis space which sabotages easy convergence towards biased hypotheses during the final optimization based on the entirety of the available data. We hereby propose a Curriculum-enhanced Group Distributionally Robust Optimization (CeGDRO) approach, which prioritizes the hardest bias-confirming samples and the easiest bias-conflicting samples, leveraging GroupDRO to balance the initial discrepancy in terms of difficulty. We benchmark our proposed method against the most popular subpopulation shift datasets, showing an increase over the state-of-the-art results across all scenarios, up to 6.2% on Waterbirds.
Introduction
Machine learning models are notorious for taking "shortcuts." In subpopulation shift scenarios, models often learn spurious correlations—such as associating a "water background" with the label "waterbird"—rather than identifying the actual semantic features of the object. This leads to catastrophic failure when the model encounters a waterbird on land at test time.
Standard Curriculum Learning (CL), which gradually increases task difficulty by starting with the easiest samples, typically exacerbates this problem. Because the "easiest" samples for a model to learn are usually those where the bias and the class align perfectly (e.g., a duck on water), learning these first "imprints" the bias into the model's weights early in the training process. Consequently, researchers in the field of robust optimization have generally avoided curriculum learning. This paper challenges that norm by asking: Can we design a curriculum that prevents bias instead of reinforcing it?
The CeGDRO Methodology
The authors propose Curriculum-enhanced Group Distributionally Robust Optimization (CeGDRO). The core philosophy is to initialize the model at an "unbiased vantage point" before the final training phase, effectively sabotaging the model's ability to converge on biased hypotheses.
The process involves five key steps:
- Environment Discovery: A simple Empirical Risk Minimization (ERM) model is trained for a single epoch to identify bias-confirming samples (likely correctly classified using shortcuts) and bias-conflicting samples (likely misclassified, requiring actual feature recognition).
- The "Anti-Curriculum" Sorting: The algorithm flips the standard CL script. For bias-confirming samples, it prioritizes the hardest ones (highest loss). For bias-conflicting samples, it prioritizes the easiest ones (lowest loss).
- Staged Training: The model is trained in stages, gradually increasing the percentage of available samples. At each stage, it draws an equal number of samples from both sorted groups.
- GroupDRO Integration: During these curriculum stages, the model uses the GroupDRO update rule to dynamically weight groups based on loss, ensuring the model stays balanced and doesn't ignore the harder bias-conflicting samples.
- Final Optimization: Once the curriculum is complete, the model undergoes a final training phase on the entire dataset.
Experimental Evaluation
CeGDRO was evaluated on three standard subpopulation shift benchmarks: Waterbirds (bird vs. background), CelebA (hair color vs. gender), and CivilComments (toxicity vs. demographic mentions). The primary metric of interest is worst-group accuracy, which measures performance on the most difficult subpopulation.
As shown in the results table, CeGDRO consistently outperformed standard GroupDRO and Invariant Risk Minimization (IRM) across all benchmarks. Most notably, on the Waterbirds dataset, CeGDRO achieved 84.8% worst-group accuracy, a significant 6.2% increase over standard GroupDRO (78.6%).
Beyond accuracy, the method demonstrated improved stability, showing lower standard deviation across multiple runs. This suggests that the curriculum makes the training process more robust and less dependent on random weight initialization.
Discussion and Conclusion
By the time a CeGDRO-trained model encounters "easy" biased data, its weights are already tuned to look for robust features, making it less likely to be swayed by spurious correlations. While the method currently relies on an initial ERM pass for environment discovery, it represents a significant step forward in integrating curriculum learning with robust optimization. Future work aims to expand this "bias-prevention" curriculum into a more general framework applicable to a wider variety of optimization protocols and complex real-world biases.
Summary updated