Quantum superposition state feature recognition method based on convolutional neural networks

In this paper, we classify cat states and coherent states by analyzing Wigner distribution images, combine CNN with the quantum state recognition task, and propose the corresponding model design and training methods. We construct both LeNet and ResNet architectures and adjust convolution kernels and strides to the optimal values. We train both LeNet and ResNet, and the loss function values indicate that ResNet performs better in classifying cat states and coherent states. We evaluate cat states and coherent states with different α values to demonstrate a certain degree of generalization capability. We added noise to the Wigner tomography with different α values. In terms of noise robustness, ResNet is more suitable for dealing with Gaussian noise and Poisson noise, and LeNet performs better in salt-and-pepper noise. The results show that LeNet may mistakenly recognize coherent states as cat states, while ResNet provides a feasible solution to solve the problem.