RetroTune: Mitigating spurious features via retrospective fine-tuning

Guangtao Zheng, Aidong Zhang

Keywords: spurious features, robustness, classification

TL;DR: We design a general self-guided spurious feature mitigation method that identifies and mitigates spurious features in the latent representation space of a model.

Abstract:

Spurious features are non-predictive and are associated with class labels in the majority of training samples. Models trained with standard empirical risk minimization tend to base their predictions on spurious features, such as identifying objects only by their frequently co-occurring backgrounds, leading to poor performance on data without the spurious features. Mitigating a model's reliance on spurious features typically requires external supervisions, such as accurate annotations of spurious features, which are not free to get.
In this paper, we propose RetroTune, a general self-guided spurious feature mitigation method that first inspects a model's latent representations based on the training samples for identifying unknown spurious features and then fine-tunes the model by targeting at the identified spurious features. Our method mimics the way of retrospection: it analyzes a model's latent representations for training samples after the model has been trained and then identifies and adjusts incorrect weights in the last classification layer of the model based on the analyzed results. RetroTune is fully unsupervised in identifying spurious features and does not need additional data to mitigate a model's reliance on spurious features. Our method achieves a maximum of 27.2\% increment in worst-group accuracy than the best baselines on training and selecting models that are robust to unknown spurious features.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7758