I was wondering what is the best way to find optimum feature map for a variational quantum circuit problem for best accuracy. I understand it’s an unexplored area. But how would one go about deciding for best feature map when presented with new dataset (like Iris,MNIST ,cancer etc). Is it a hit and trial process or is there a systematic way to do it.
Hi @P.Kairon! This is quite a new area of research, and as far as I am aware there is no systematic approach to determining an optimum feature map (@Maria_Schuld may know of some potential heuristics or tips and tricks here).
Something that might be of interest, however, is our demo on embeddings and metric learning based on the paper https://arxiv.org/abs/2001.03622. In this demo, we train the embedding/feature map in order to maximize separation of features in the quantum Hilbert space.
Josh is right, the only thing I wanted to add is another demo which might interest you.
Here you can see that common embeddings create quantum models that look like drastically truncated Fourier series.