17_gen_2017

Harmonic Networks: Deep Translation and Rotation Equivariance

Questions:

[gitabcworld]: Equivariance condition is very important in H-Net. The summation of two responses of the same order m must reamin of order m. They enforce this property at each feature map. They say they combine only feature maps of the same rotation order. How do they control this property in the implementation in the H-Net?

[prlz77]: Regarding the previous question (Albert), I think the streams disentagle de different rotation orders, and then they are all unified in the end. Am I wrong?

[@priba]: "Some deep learning libraries implement cross-correlation rather than convolution". Which are the benefits of using cross-correlation in a deep learning library? (Basic question) How it deal with the problems of 9 and 6?