NeurIPS 2020

Learning to Adapt to Evolving Domains


Meta Review

After the rebuttal and discussion phase, three reviewers are leaning marginally positive, while reviewer #9 still has concerns. The authors claim that they are studying a new setting which assumes that the target domain is changing rapidly and data from target is arriving in small batches and cannot be stored. R9 is concerned that "(1) the motivation behind proposing the evolving domain adaptation setting, compared to other similar settings (e.g. continuous DA), was unclear, (2) the comparison with baseline methods was not comprehensive - many existing methods can be adapted to the proposed problem with minor modifications." The authors provided a rebuttal where they explain the difference between related work on continuous DA and their setting and provide additional comparisons to existing methods. The rebuttal did not convince R9. After reading the paper and rebuttal, the AC agrees with the authors that there is a difference between their problem setting and other continuous DA work, which is the requirement to avoid catastrophic forgetting on older data as the model adapts to the new data. The AC therefore thinks that the paper can contribute something new to the existing body of research by studying this new problem. The authors should explain this setting difference more clearly in the paper, especially in Figure 1. The authors are also encouraged to include the additional experiments comparing to existing methods for continuous DA in the camera ready.