A direct error correction method for quantum machine learning

In performing quantum machine learning on quantum computers (noisy intermediate-scale quantum devices), we need quantum noise and error mitigation methods to obtain reliable outcomes. We thus propose a correction method using learning-based error mitigation for hybrid quantum-classical machine learning, which can be used by experimenters without prior knowledge of the quantum noise channel models. A correction function using linear regression models inspired by the depolarizing noise channel is trained, and the function is to correct classical probabilities of each qubit of the quantum circuit for the quantum machine learning model. In constructing the training data for the correction function, we assume that quantum circuit simulators yield classical probabilities with the noise-free state as the ground truth. In the experiments, quantum generative models, quantum neural networks, and quantum circuit learning models were used for demonstrating the effectiveness of the proposed method on noisy quantum devices (simulation models) and real devices. Experimental results demonstrated that the proposed method worked well for the error correction, and the convergence speed of training the models was improved.