Speaker
Description
Neural networks provide a powerful framework for learning quantum states directly from measurement data across extended regions of parameter space. However, acquiring the data required to train these models can represent a significant cost. In this work, we introduce an active learning strategy for improving data acquisition for training conditional generative models. Based on a query-by-committee protocol, the strategy iteratively acquires data in phase space regions where an ensemble of models exhibits larger disagreement in sample distribution. The resulting trained models can identify phase transitions through susceptibility-based analyses of the learned distributions, without requiring prior knowledge of the relevant order parameters. We show that to achieve a comparable performance with uniform sampling, active learning requires approximately $3\times$ less data, representing a considerable saving in compute cost and acquisition time.
| Topic | Machine learning techniques for experimental quantum information science |
|---|