I have not investigated this as much as I should have but if I remember correctly there were cases where the NN approach was yielding smaller solutions. Would b

I have not investigated this as much as I should have but if I remember correctly there were cases where the NN approach was yielding smaller solutions. Would be a great follow up to the thesis. The way this was turned into an optimization problem was to assume a NN to input an identity matrix and then have a custom layer in there to generate S(2) unitaries (exponential form) for which the phase parameters are then the learned parameters in the NN. Similarly for global XX gates. Then from this the final unitary can be computed and compared against the desired unitary and a loss function can be derived. I remember it was a little fiddly to implement these custom layers in tensorflow since many of the functions didn’t work for imaginary numbers. But yeah in short a circuit structure was assumed (alternating between a global XX and single qubit operations) for which the phases of their generating matrices were the learned parameters of the network. Then multiple topologies (how many steps in the QC) could be validated. I think the coolest result was that NNs consistently outperformed other optimisation strategies to learn those parameters.