Speaker
Description
With the increasing size of NISQ devices, it is important to find novel approaches to classically simulate large quantum circuits. Recent works show that machine learning (ML) models allow the efficient simulation of variational quantum algorithms (like QAOA). However, although ML models typically achieve great success in simulating quantum circuits, there are cases where such models perform worse than expected.
In this talk, we will try to answer three questions: why does ML underperform? When does it underperform? And, how can we improve the simulation fidelity? We will show that the simulation quality is highly related to the sample quality and simulation path. We will also show that a circuit can be difficult to simulate when the entanglement entropy of the target state is high. Finally, to mitigate the quantum state approximation error, we proposed two heuristics: an MCMC sampler that utilizes the problem symmetry and a greedy strategy to choose the simulation path.
