Why use complex number in convolution

[jjjjohnson]

Hi @denfed
The paper says "we compute instead the convolution with 2N real-valued filters φ ̃ , n = 1, . . . , 2N , and perform squared l2-pooling". I am wandering why in the code we still use complex number when initializing the filter.

Thanks for the great work!
Junjie

[denfed]

Hello! My apologies for the extremely late reply on this. The authors of LEAF use 2N real-valued filters, each for the real and complex portions of complex numbers. It's a pretty cool trick for an easy way of working with complex numbers without actually having to use complex tensors for convolutional layers. For example, 2+4i and 3+2i would be represented as two real valued filters as [2,3] and [4,2].

When we initialize the filters, we do create complex numbers but we right away split them into two real-valued filters, shown above.