- Unsupervised Idea Discovery Mitigates Spurious Correlations(arXiv)
Authors: Md Rifat Arefin, Yan Zhang, Aristide Baratin, Francesco Locatello, Irina Rish, Dianbo Liu, Kenji Kawaguchi
Summary: Fashions susceptible to spurious correlations in coaching knowledge typically produce brittle predictions and introduce unintended biases. Addressing this problem sometimes includes strategies counting on prior data and group annotation to take away spurious correlations, which will not be available in lots of purposes. On this paper, we set up a novel connection between unsupervised object-centric studying and mitigation of spurious correlations. As a substitute of straight inferring sub-groups with various correlations with labels, our method focuses on discovering ideas: discrete concepts which can be shared throughout enter samples. Leveraging present object-centric illustration studying, we introduce CoBalT: an idea balancing approach that successfully mitigates spurious correlations with out requiring human labeling of subgroups. Analysis throughout the Waterbirds, CelebA and ImageNet-9 benchmark datasets for subpopulation shifts reveal superior or aggressive efficiency in contrast state-of-the-art baselines, with out the necessity for group annotation
