How are circuit architectures that work well found? e.g Quantum transfer learning — PennyLane
I am currently investigating a hybrid neural network for the MNIST dataset. What would be a good approach to find a good circuit architecture that works?
Thanks for the help
This is still a bit of an art and an open question research-wise. Since it’s still early for QML, the community hasn’t really settled upon any “canonical” circuit architectures yet.
People either create their own (more or less arbitrarily in some cases), or just duplicate an architecture they’ve seen in other papers. That’s usually a good starting point. Beyond that, you will likely have to experiment a bit with different architectures, to find the one that seems to work best for your problem.
In the latest master branch of PennyLane (not yet released) we have some new general functionality for creating circuit architectures (e.g., the
broadcast function), along with pre-existing template circuits (https://pennylane.readthedocs.io/en/latest/introduction/templates.html). I encourage you to check those out as you’re exploring: they could prove helpful!