New Method Identifies Spurious Correlations in ML Models.

Machine learning models trained with empirical risk minimization often achieve high average accuracy but fail on specific subpopulations due to spurious correlations in the training data. Existing methods to identify these problematic samples typically require group annotations or signals from early training, which can be difficult to implement or require careful hyperparameter tuning. This paper introduces a simpler, more robust signal for identifying spuriously correlated samples, available after a model has converged. At convergence, while the overall loss may no longer differentiate populations, samples consistent with spurious correlations are classified by a shared, robust rule. In contrast, samples that defy the spurious correlation are fit through input-specific configurations, making their predictions more fragile. The proposed procedure involves applying a fixed perturbation to a converged model's inputs. It has been observed that predictions for the spuriously correlated samples are far less likely to flip than those for the fragile, non-spuriously correlated samples. This method requires only two forward passes per training sample, no group annotations at any stage, and no early-stopping epoch. Using the detected samples to rebalance training data significantly improves worst-group accuracy, for example, raising it from 57.3% to 80.8% on the Waterbirds dataset, approaching the performance achieved with ground-truth group labels.