Have unitary/orthogonal matrices or circulant/Fourier matrices (e.g. constructed by repeating initial input samples X) been tried as weight matrices W for Deep Learning - where a smaller set of parameters such as angles/phases is tuned?
Hi, @Sumeet_Sandhu.
It’s not a topic I’m familiar with, and it’s not discussed anywhere in this specialization, but you may find EUNN, OrthDNN and CirCNN interesting.
Feel free to share any insights here 