Southwest Quantum Information and Technology

Error mitigation of superconducting qubits in generative modeling benchmark tasks

Presenting Author: Kathleen Hamilton, Oak Ridge National Laboratory
Contributing Author(s): Raphael C. Pooser

Application-based benchmarks for near-term quantum devices rely on multi-component computational tasks to evaluate hardware capabilities. We have used the gradient-based training of parameterized circuit models to quantify the performance of noisy superconducting qubits for generative modeling tasks [1]. A classical gradient-based optimizer (e.g. Adam) is used to train the circuit parameters and the quantum hardware is used to evaluate the loss function gradient. The difficulty of this benchmark is partially due to the number of noisy quantum circuit executions required to evaluate the gradient function. The incorporation of error mitigation into a gradient-based training workflow is a non-trivial task. Error mitigation has the potential to speed up gradient-based training by reducing spurious information passed to a classical optimization method; but the loss of relevant information can impair training. In this talk we will present results for matrix-based error mitigation used inside gradient-based training of circuits trained on superconducting qubits, highlighting specific use cases and discussing several challenges. [1] Hamilton, Kathleen E., Eugene F. Dumitrescu, and Raphael C. Pooser. "Generative model benchmarks for superconducting qubits." Physical Review A 99.6 (2019): 062323. This work was supported as part of the ASCR Testbed Pathfinder Program at Oak Ridge National Laboratory under FWP #ERKJ332

(Session 5 : Saturday from 5:00pm - 7:00pm)