[논문 리뷰] MaskTune: Mitigating Spurious Correlations by Forcing to Explore

Saeid Asgari Taghanaki, Aliasghar Khani외 5명|arXiv (Cornell University)|2022. 09. 30.

Machine Learning and Data Classification인용 수 20

한 줄 요약

MaskTune은 trained model이 발견한 가장 판별력이 높은 특징을 마스킹하여 새로운 입력 특징의 탐색을 강제하고, 감독 없이도 허위 상관관계에 대한 의존도를 줄이는 단일-에폭 파인튜닝 방법이다.

ABSTRACT

A fundamental challenge of over-parameterized deep learning models is learning meaningful data representations that yield good performance on a downstream task without over-fitting spurious input features. This work proposes MaskTune, a masking strategy that prevents over-reliance on spurious (or a limited number of) features. MaskTune forces the trained model to explore new features during a single epoch finetuning by masking previously discovered features. MaskTune, unlike earlier approaches for mitigating shortcut learning, does not require any supervision, such as annotating spurious features or labels for subgroup samples in a dataset. Our empirical results on biased MNIST, CelebA, Waterbirds, and ImagenNet-9L datasets show that MaskTune is effective on tasks that often suffer from the existence of spurious correlations. Finally, we show that MaskTune outperforms or achieves similar performance to the competing methods when applied to the selective classification (classification with rejection option) task. Code for MaskTune is available at https://github.com/aliasgharkhani/Masktune.

연구 동기 및 목표

• 주석이 달린 허위 레이블을 사용하지 않고 심층 모델에서 잘못된 입력 특징에 과도하게 의존하는 것을 동기 부여하고 해결한다.
• 추가 특징의 탐색을 강제하는 마스킹 기반 파인튜닝 기법을 제안한다.
• 다양한 데이터셋에서 허위 상관관계에 대한 강건성과 선택적 분류에의 적용 가능성을 입증한다.
• 마스킹이 모델 복잡도와 동작에 어떻게 영향을 미치는지에 대한 경험적 및 이론적 통찰을 제시한다.

제안 방법

• Train an ERM model on original data.
• Compute a feature importance map with xGradCAM for each training sample.
• Mask the most discriminative regions using a thresholded map and up-sampling (masking the top contributions).
• Finetune the fully trained model for one epoch on the masked dataset with a small learning rate.
• Repeat with masking to encourage discovery of new complementary features.
• Optionally ensemble the original and MaskTune models for selective classification.

실험 결과

연구 질문

• RQ1 Can masking the initially discovered discriminative features during a single-epoch finetune reduce reliance on spurious correlations without supervision?
• RQ2 Does MaskTune encourage learning a broader set of input features, improving robustness and enabling reliable abstention in selective classification?
• RQ3 How does MaskTune perform across datasets with different spurious correlations and in selective classification scenarios?

주요 결과

• MaskTune outperforms or matches competing methods on biased MNIST, CelebA, Waterbirds, and Background Challenge datasets.
• MaskTune yields better worst-group accuracy on CelebA and Waterbirds without group supervision while maintaining high overall accuracy.
• MaskTune improves selective classification performance, achieving strong results on CIFAR-10, SVHN, and Cats vs. Dogs across coverage levels.
• The method does not require spurious feature annotations or subgroup labels during training or model selection.
• MaskTune increases reliance on a broader set of input features, as visualized by explainability maps.
• Theoretical considerations show masking can increase model complexity and force exploration in over-parameterized regimes.