The reliability of this method. about club

[Linear95]

Based on the theorem in our paper, the performance of CLUB is based on the quality of the variational approximation q_theta(y|x). Since the dimension of the sample (x,y) is changed, the variational approximation q_\theta(y|x) is supposed to be changed to have a better estimation. My suggestion is to reduce the hidden_size of q_\theta(y|x) for a better approximation to data samples with lower dimensions.

[LindgeW]

@erow InfoNCE不是互信息最大化的近似估计吗？为什么和CLUB比较？

[erow]

@erow InfoNCE不是互信息最大化的近似估计吗？为什么和CLUB比较？

目的都是为了估计互信息，只是有的方法是最大化求MI的下界，有的是最小化求MI的上界。