NeurIPS 2020

Interventional Few-Shot Learning

Meta Review

This paper analyzes few-shot learning from a causal inference perspective and presents an interesting claim that pretrained knowledge is a confounder that limits performance. The authors use this finding to propose an interventional few shot learning paradigm. I think this is a solid paper where the theoretical insight results in a good empirical performance. R2 has a serious concern regarding the clarity of the paper that makes it difficult to verify the correctness of the paper, in particular with respect to the kind of confounding effect that this approach can resolve. I think this is a valid concern and I suggest the authors attempt to fully address this in the final version of the paper.