Beyond Gates

Coefficients in quantum Fourier models are determined (with some constraints) by the parameters of the trainable unitaries. Naturally the degree to which coefficients can be adjusted is given by the number of paramters available to the unitaries (Note: this is a simplification). Now, in almost every (non-trivial) training scenario, areas in the loss landscape exist where the gradient becomes very small but the global optimum is not reached. These points are commonly referred to as local minima (in case the area is convex) or critical points (in case it's a saddle point, so the local area has a gradient). Both scenarios may slow down optimization or bring it to a halt in the worst case.

With our recent changes to the [QML-Essentials Framework], we've introduced [Pulse Level Parameters] to our quantum Fourier models. Given this setting, we've shown that access to pulse parameters consistingly helps the optimization to achieve better results. These results even surpass the ones achieved through overparameterization, a concept that is in principle very familiar but cannot fully cover the dynamics introduced by changing pulse parameters. This work is currently under review, but you can find a preprint on Arxiv.