Speaker
Description
Quantum error mitigation (QEM) is a near-term route to reducing bias in estimates extracted from noisy quantum processing units (QPUs). We present a systematic study of machine learning (ML)-guided QEM: deep learning models that mitigate full output probability distributions from measured quantum circuits. The task is deliberately stringent: an improved computational-basis distribution can be used to estimate any diagonal observable, but it requires the model to learn correlations across all bitstrings rather than a single expectation value. We compare fully connected, recurrent, Transformer, noisy-encoder, and Perceiver-style architectures; contrast direct prediction models with multiplicative correction models; and test pre-training, fine-tuning, ablation, and transfer-learning workflows. On noisy simulations parameterised by IBM calibration data and on real data from IBM superconducting QPUs, attention-based architectures -- especially Transformer and Perceiver variants -- are the most reliable learned models; they are competitive with strong analytic baselines on simulated data, outperform the tested post-processing baselines on real hardware data, and transfer effectively between comparable devices.
| Topic | Machine learning techniques for experimental quantum information science |
|---|