Title:	Low Sample Quantum Machine Learning Simulation
DNr:	LiU-compute-2024-22
Project Type:	LiU Compute
Principal Investigator:	Nathaniel Helgesen <nathaniel.helgesen@liu.se>
Affiliation:	Linköpings universitet
Duration:	2024-05-28 – 2025-06-01
Classification:	10299
Keywords:

Abstract

The field of quantum machine learning seeks to reach quantum supremacy in machine learning, but is constantly stifled by necessarily high sampling needs caused by noise in quantum computers and statistical bounds on quality estimates of expectation values. While many work to reduce the noise in quantum computations through both better hardware and error correction, less work is put towards increasing the confidence of VQCs and decrease the sampling needs of the algorithms themselves. Many proven uses of quantum computers require floating point values, and thus impose a high sampling rate in order to estimate the uncollapsed internal state of the quantum computer, but even classification, which can make use of discrete quantum outputs, has been ignored as a potential avenue to study ways to improve model confidence and decrease sampling needs. In this paper, we focus on multiclass classification and describe a parameter-free post-processing technique that treats each wire as an independent binary decision. This method improves few-sample accuracy during inference by disentangling the wire outputs and forcing the VQC to avoid uncertain outputs. We provide a description of this method as well as several metrics of accuracy, entanglement, and gradient magnitude in order to make a push towards using few-sample accuracy as a primary goal for effective VQCs.