NeurIPS 2020

Heuristic Domain Adaptation

Meta Review

The paper addresses domain adaptation from the perspective of heuristic search, modeling explicitly the domain-specific characteristics. Transferrable representations were extracted from a fundamental network and a heuristic network, with the latter decomposed into several subnetworks. Three constraints were proposed on the transferrable representations: similarity, independence, and termination. The method is shown effective on challenging datasets such as office-home, with superior performance over several state of the art. Overall, this paper provides an interesting idea using heuristic search. The reviewers raised several concerns on experiments including ablation study. They have been addressed well by the rebuttal, and can be incorporated to the final version of the paper. I think this novel approach deserves publication at NeurIPS.