NeurIPS 2020

Learning from Failure: De-biasing Classifier from Biased Classifier

Meta Review

The paper makes an interesting observation that cometimes, the classifier "biases" can be detected from examples that the models corectly classify early on. They present the first successful unsupervised method to correct for such biases, with promising experimental results since it compares reasonably compared to supervised approaches on artificial and real datasets. Moreover, the reviewers noted that the paper tackles a very timely and important topic.