A Conditional Generative Model Based on Quantum Circuit and Classical Optimization

Generative model is an important branch of unsupervised learning techniques in machine learning. Current research shows that quantum circuits can be used to implement simple generative models. In this paper, we train a quantum conditional generator, which can generate different probability distributions according to different input labels, i.e., different initial quantum states. The model is evaluated with different datasets including chessboard images, and bars and stripes (BAS) images of 2 x 2 and 3 x 3 pixels. We also improve the performance of the model by introducing a controlled-NOT (CNOT) layer. The simulation results show that the CNOT layer can improve the performance, especially for the generative model with chain-connected entangling layers.