Is Last Layer Re-Training Truly Sufficient for Robustness to Spurious Correlations?

Last-layer retraining alone is insufficient: While the Deep Feature Reweighting (DFR) method improves performance on underrepresented groups, the paper finds that it "remains susceptible to spurious correlations," meaning models can still rely on undesired auxiliary features even after this targeted fix.

The assumption about "core features" may not hold in practice: DFR is based on the premise that standard models learn useful core features well enough that only the final layer needs adjustment. This paper challenges that assumption, particularly in realistic medical imaging applications where the underlying data is more complex.

Medical domain presents unique challenges: The researchers specifically tested DFR on medical data and found it less applicable than in controlled settings, suggesting that real-world healthcare applications require more robust solutions than simple last-layer retraining to address bias and spurious correlations.

Worst-performing groups remain vulnerable: Even with DFR applied, the paper demonstrates that certain subgroups of patients—those missing expected features or having unexpected feature combinations—continue to experience degraded accuracy, which has serious implications for equitable AI in healthcare.

More comprehensive approaches are needed: The findings suggest that addressing spurious correlations requires deeper intervention than retraining final layers, pointing toward the need for more fundamental changes to how models are trained or validated to ensure fairness across all demographic groups.