Abstract: In this report, we attempt to verify the claims that the paper makes about their proposed CGN framework that decomposes the image generation process into independent causal mechanisms. Further, the author claims that these counterfactual images improves the out-of-distribution robustness of the classifier. We use the code provided by the authors to replicate several experiments in the original paper and draw conclusions based on these results.
Paper Url: https://arxiv.org/abs/2101.06046
Paper Venue: ICLR 2021