Unsupervised Learning of Unbiased Visual Representations

Abstract

Deep neural networks are known for their inability to learn robust representations when biases exist in the dataset. This results in a poor generalization to unbiased datasets, as the predictions strongly rely on peripheral and confounding factors, which are erroneously learned by the network. Many existing works deal with this issue by either employing an explicit supervision on the bias attributes, or assuming prior knowledge about the bias. In this work we study this problem in a more difficult scenario, in which no explicit annotation about the bias is available, and without any prior knowledge about its nature. We propose a fully unsupervised debiasing framework, consisting of three steps: first, we exploit the natural preference for learning malignant biases, obtaining a bias-capturing model; then, we perform a pseudo-labelling step to obtain bias labels; finally we employ state-of-the-art supervised debiasing techniques to obtain an unbiased model. We also propose a theoretical framework to assess the biasness of a model, and provide a detailed analysis on how biases affect the training of neural networks. We perform experiments on synthetic and real-world datasets, showing that our method achieves state-of-the-art performance in a variety of settings, sometimes even higher than fully supervised debiasing approaches.