NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:1291
Title:Information Competing Process for Learning Diversified Representations


		
This paper received two positive reviews and one quite negative review. So I read it carefully. The rebuttal has addressed the third point of Reviewer #2. I agree that maximizing I(z,t|\phi_z) + I(y,t|\phi_y) is sufficient for maximizing I(r,t|\phi), but this worry is more about redundancy than detriment. I personally don't find multiview so relevant here, although the technique of minimizing the mutual information between z and y is not new. Variational approximation is used extensively, and it looks effective. Most reviewers raised ablation study, but I think the paper has already conducted the most important ones: ICP-ALL and ICP-COM. Overall, the paper puts together some mutual information terms to learn diversified representations, facilitated by variational approximation. The experimental results look quite encouraging. So I would recommend it for publication. In the camera-ready, please address point e) of Reviewer #1, and incorporate all the modifications promised in the rebuttal.