Loading...

Robust Learning to Spurious Correlation without Access to Side Information of the Environment

Ghaznavi, Mahdi | 2025

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 58193 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Rohban, Mohammad Hossein; Soleymani Baghshah, Mahdieh
  7. Abstract:
  8. Traditionally, machine learning models for classification tasks rely on statistical methods to find correlations between patterns in the input data and their correspond- ing labels. However, these correlations are not necessarily consistent across different data partitions and may change at test time. Such unstable correlations are referred to as spurious correlations. When the spurious correlation relied upon during training changes at test time, the model’s accuracy can degrade. To improve robustness to shifts in spurious correlations, most research in this area assumes that group annota- tions based on different values of the spurious attribute are available during training or validation. In this context, a ”group” is defined as a set of samples with the same label and a particular spurious attribute. However, when group labels are unavailable or the spurious correlations are unknown, existing methods become inapplicable. This research introduces Environment-based Validation and Loss-based Sampling (EVaLS) as a fully annotation-free approach to overcome these limitations. The method identi- fies high- and low-loss samples from a model trained using Empirical Risk Minimization (ERM) in order to construct a balanced dataset. This approach improves robustness to correlation shift in 3 out of 5 benchmarks compared to peer methods with group super- vision. Then, using environment inference, it constructs environments with correlation shifts, without requiring any labeled validation data. By using worst-environment ac- curacy as a surrogate for worst-group accuracy in model selection, EVaLS guides the retraining of the last linear layer of the pre-trained model, leading to improved ro- bustness against spurious correlations. The proposed approach achieves competitive robustness compared to methods that require group labels. Moreover, the absence of any need for group labels at any stage enables EVaLS to handle unknown spurious correlations—a key advantage over existing methods. This makes EVaLS particularly suitable for real-world scenarios involving complex and multiple spurious correlations. This is demonstrated by a 15.9% increase in the worst group accuracy of the method proposed in this thesis compared to the best known method that relies on spurious correlation
  9. Keywords:
  10. Out-Of-Distribution Generalization ; Robustness ; Spurious Correlation ; Causality ; Unknown Spurious Correlations ; Subpopulation Shift ; Invariant Learning

 Digital Object List

 Bookmark

...see more