Deep learning models often rely only on a small set of features even when there is a rich set of predictive signals in the training data. This makes models brittle and sensitive to distribution shifts. In this work, we first examine vision transformers (ViTs) and find that they tend to extract robust and spurious features with distinct attention heads. As a result of this modularity, their performance under distribution shifts can be significantly improved at test time by pruning heads corresponding to spurious features, which we demonstrate using an "oracle selection" on validation data. Second, we propose a method to further enhance the diversity and complementarity of the learned features by encouraging orthogonality of the attention heads' input gradients. We observe improved out-of-distribution performance on diagnostic benchmarks (MNIST-CIFAR, Waterbirds) as a consequence of the enhanced diversity of features and the pruning of undesirable heads.
View on arXiv