LWKJJONAK
commented
1 year ago Actually, the classification machine is deployed on the quantum computer. The quantum device can be utilized to deploy the QNN model and calculate the gradients, then the model parameters can be optimized with these gradients (optimizers such as Adam can be utilized as well). Let me use a more concrete example:
In our recent work Experimental Quantum Adversarial Learning with Programmable Superconducting Qubits (accepted by Nature Computational Science), we implemented such a QNN classifier on a 36-qubit superconducting quantum processor. In this work, we deployed both the block-encode QNN (for classical medical images) and amplitude-encode QNN (for quantum many-body state data) on quantum devices. With the parameter shift rule, we can calculate the gradient of the loss function w.r.t the model parameters, and further update these parameters. The training loop is executed on this quantum processor (i.e. these are not numerical simulations on our PC but are real quantum experiments) (of course, control software is utilized to monitor the process and collect data).
erlebach
I have a question on the Amplitude-encoding demo. On a real Quantum Device, what parts of the simulation are done on the quantum computer and what part is done on a classification machine? For example, where is the training loop executed? Thanks.