Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·Aug 7·5 min read
Perturbation Sensitivity at Convergence: A Simple Signal for Identifying Spuriously Correlated Samples
TL;DR
A new method identifies spurious correlations in machine learning models post-convergence by measuring perturbation sensitivity, improving accuracy on challenging subpopulations. This approach boosts worst-group accuracy on the Waterbirds dataset from 57.3% to 80.8% without requiring group annotations.
✦ Why It Matters
Implement this perturbation sensitivity method to improve model robustness against spurious correlations in your datasets.
Key Takeaways
- Use perturbation sensitivity to identify fragile samples in your model.
- No group annotations are needed for this method.
- Apply this technique post-convergence for better accuracy.
- Consider rebalancing training data based on identified samples.
How It Works
The method assesses perturbation sensitivity by applying fixed perturbations to model inputs after convergence. Samples that conform to spurious correlations exhibit consistent predictions, while those that do not are more sensitive to input changes, leading to prediction flips.
This distinction allows for effective identification of problematic samples without early stopping or group labels.