Towards Optimal Quantum Neural Network Structure
Yadong Wu (吴亚东) 博士 清华大学
摘要：Machine learning based on neural networks has recently provided significant advances for many practical applications. In physics, one natural application is the study of the quantum many-body systems. There are several works about using quantum computation or quantum neural networks to enhance conventional machine learning tasks. Intuitively, for the deep learning, we thought the more complicated the neural networks, better the expressions. But we can’t increase the depth of the neural network infinitely. So here is the question, for the finite depth of a neural network, or more specifically, for the same two-bit gate number of a quantum neural network, what kinds of circuit structure can give us a good prediction after training. In this work, we use two qualities of a quantum circuit to describe the expression ability of a quantum circuit structure. One is the tripartite information, other is the operator size. We build a few different kinds of quantum circuit structures and compare the learning ability of these structures. After learning a quantum task and a classical task, we find the more scramble of a circuit, the better it can learn. What’s more, we design ‘the super circuit’, which has the best learning ability.