BAdd: Bias mitigation through bias addition

Ioannis Sarridis, Christos Koutlis, Symeon Papadopoulos, Christos Diou

Abstract: Computer vision datasets often exhibit biases in the form of spurious correlations between certain attributes and target variables. While recent efforts aim to mitigate such biases and foster bias-neutral representations, they fail in complex real-world scenarios. In particular, existing methods excel in controlled experiments on benchmarks with single-attribute injected biases, but struggle with complex multi-attribute biases that naturally occur in established CV datasets. In this paper, we introduce BAdd, a simple yet effective method that allows for learning bias-neutral representations invariant to bias-inducing attributes. This is achieved by injecting features encoding these attributes into the training process. BAdd is evaluated on seven benchmarks and exhibits competitive performance, surpassing state-of-the-art methods on both single-and multi-attribute bias settings. Notably, it achieves+ 27.5% and+ 5.5% absolute accuracy improvements on the challenging multi-attribute benchmarks, FB-Biased-MNIST and CelebA, respectively.