A hybrid quantum-classical generative adversarial networks algorithm based on inherited layerwise learning with circle-connectivity circuit

Quantum generative adversarial networks (QGANs) have a potential exponential advantage over classical GANs, which has attracted widespread attention. However, it also faces the barren plateau problem, i.e., the gradient of variational quantum circuits (VQCs) initialized with random initial parameters decreases exponentially as the number of circuit layers and parameters increases. In order to solve this problem, a hybrid quantum-classical GANs algorithm based on inherited layerwise learning with circle-connectivity circuit (ILL-QGAN) is proposed. The shallow-depth circuit is gradually added during optimization process, and only subsets of parameters are updated, which reduces the number of parameters and circuit depth in each training step. Besides, the parameters trained from the previous layer are inherited to the next layer, which provides the favorable initial parameters for the later. In addition, in order to offer relatively favorable expressibility and entangling capability, the generator uses a more near-term circuit structure, i.e., circle connectivity, to construct VQCs. Finally, the BAS and handwritten experiments are conducted on PennyLane to show our algorithm can faster converge to the Nash equilibrium point and obtain higher accuracy than existing quantum algorithms. Since the shallow-depth and circle-connectivity circuit are used, our algorithm is preferable for execution on noisy intermediate-scale quantum devices.