Removing Spurious Concepts from Neural Network Representations via Joint Subspace Estimation

Out-of-distribution generalization in neural networks is often hampered by spurious correlations. A common strategy is to mitigate this by removing spurious concepts from the neural network representation of the data. Existing concept-removal methods tend to be overzealous by inadvertently eliminating features associated with the main task of the model, thereby harming model performance. We propose an iterative algorithm that separates spurious from main-task concepts by jointly identifying two low-dimensional orthogonal subspaces in the neural network representation. We evaluate the algorithm on benchmark datasets for computer vision (Waterbirds, CelebA) and natural language processing (MultiNLI), and show that it outperforms existing concept removal methods