Sampling based quantum training of arbitrary spin-graphs

Machine Learning algorithms trainable on a quantum device are fast emerging as

a robust alternative to the classical ones in a wide variety of tasks as they exploit

the power of quantum devices. Of particular interest are spin-graphs which has been

recently proven to offer an expressive and efficient neural-network based representation

of quantum states of a given system.1–3 In this work, we show that arbitrary such

graphs can be trained on a quantum device using a sampling technique that is capable of

demonstrating quantum advantage. Besides, the algorithm is linear scaling with system

size in qubit resources and gate-requirements. To demonstrate its efficacy, we exemplify

the protocol on a wide-variety systems ranging from strongly correlated molecules to

important class of spin models which forms the basis of quantum magnetism in 1D and

2D. Apart from such wide trainability, the results are in reasonable agreement with

those obtained from conventional means which serves to explicate that our algorithm

can offer a promising alternative to traditional standards even using noisy near-term

devices.