Simple Disentanglement of Style and Content in Visual Representations
This work addresses the challenge of disentangling features for large-scale vision datasets, offering a practical solution for domain generalization in scenarios with style changes or spurious correlations.
The paper tackles the problem of learning interpretable visual representations by proposing a simple post-processing framework to disentangle content and style in pre-trained vision model features, resulting in significant domain generalization improvements for style-related distribution shifts.
Learning visual representations with interpretable features, i.e., disentangled representations, remains a challenging problem. Existing methods demonstrate some success but are hard to apply to large-scale vision datasets like ImageNet. In this work, we propose a simple post-processing framework to disentangle content and style in learned representations from pre-trained vision models. We model the pre-trained features probabilistically as linearly entangled combinations of the latent content and style factors and develop a simple disentanglement algorithm based on the probabilistic model. We show that the method provably disentangles content and style features and verify its efficacy empirically. Our post-processed features yield significant domain generalization performance improvements when the distribution shift occurs due to style changes or style-related spurious correlations.