Revealing the working mechanism of quantum neural networks by mutual information

TL;DR

This paper investigates the training mechanism of quantum neural networks using mutual information, revealing how information flows during training and correlates with learning and generalization.

Contribution

It introduces a method to analyze QNNs by dividing the circuit into subsystems and tracking mutual information, providing insights into their working mechanism.

Findings

Mutual information I(Di : Mo) increases during training, indicating information transfer to measurement subsystem.

I(Mi : Mo) exhibits a two-phase behavior, first increasing then decreasing, aligning with feature fitting and generalization.

The analysis links mutual information dynamics to QNNs' accuracy and generalization capabilities.

Abstract

Quantum neural networks (QNNs) is a parameterized quantum circuit model, which can be trained by gradient-based optimizer, can be used for supervised learning, regression tasks, combinatorial optimization, etc. Although many works have demonstrated that QNNs have better learnability, generalizability, etc. compared to classical neural networks. However, as with classical neural networks, we still can't explain their working mechanism well. In this paper, we reveal the training mechanism of QNNs by mutual information. Unlike traditional mutual information in neural networks, due to quantum computing remains information conserved, the mutual information is trivial of the input and output of U operator. In our work, in order to observe the change of mutual information during training, we divide the quantum circuit (U operator) into two subsystems, discard subsystem (D) and measurement subsystem (M) respectively. We calculate two mutual information, I(Di : Mo) and I(Mi : Mo) (i and o means input or output of the corresponding subsystem), and observe their behavior during training. As the epochs increases, I(Di : Mo) gradually increases, this may means some information of discard subsystem is continuously pushed into the measurement subsystem, the information should be label-related. What's more, I(Mi : Mo) exist two-phase behavior in training process, this consistent with the information bottleneck anticipation. The first phase, I(Mi : Mo) is increasing, this means the measurement subsystem perform feature fitting. The second phase, I(Mi : Mo) is decreasing, this may means the system is generalizing, the measurement subsystem discard label-irrelevant information into the discard subsystem as many as possible. Our work discussed the working mechanism of QNNs by mutual information, further, it can be used to analyze the accuracy and generalization of QNNs.