Readout Noise Mitigation with Bayesian Methods for Quantum Neural Networks

ICLR 2026 Conference Submission20966 Authors

Keywords: Readout noise, noise mitigation, bayesian method, quantum neural network

Abstract: Readout noise remains a significant barrier for variational quantum circuits (VQCs) and quantum neural networks (QNNs), as incorrect observations modify gradients, bias optimization, and lower predictive accuracy. For readout mitigation, we use Bayesian inference and suggest a two-step approach that takes drift into account. Offline, a Bayesian neural network (BNN) converts noisy shot histograms into corrected outcome distributions with calibrated uncertainty, resulting in expressive, data-driven priors. Online, iterative Bayesian unfolding (IBU) starts with these priors and updates with current calibration counts; an uncertainty-adaptive stopped rule prevents overfitting to temporary drift. Experiments on CIFAR-10 and EuroSAT, chosen to demonstrate robustness across both vision and remote-sensing domains, show that our method achieves up to 12.4\% reduction in classification error and 9.8\% improvement in training stability compared to confusion-matrix correction and ML-based baselines such as logistic regression, shallow neural networks, and probabilistic noise models. Importantly, uncertainty-adaptive iteration control enables our framework to balance offline priors with fresh observations, preventing overfitting to noise. Beyond quantum applications, this illustrates a general learning principle: Bayesian priors combined with online refinement offer a scalable path toward robust learning under dynamic and nonstationary noise.

Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)

Submission Number: 20966