Quantum Learning from Label Proportion

Yu Yamaoka, Weng Ian Chan, Kosuke Mitarai

Keywords: Quantum machine learning, Learning from label proportion

TL;DR: A quantum version of weakly supervised LLP that use probabilistic nature of quantum measurements.

Abstract:

Learning from Label Proportions (LLP) is a weakly supervised learning method in which training data are provided as bags of instances annotated only with class proportions.
We introduce Q-LLP, a quantum formulation of LLP, which directly uses probabilistic measurement arising from superposition in quantum computers.
Quantum machine learning is theoretically expected not only to enhance computational power in general but also to prevent overfitting and improve representation by processing capabilities in high-dimensional Hilbert spaces.
However, executing conventional learning methods on quantum computers requires algorithmic translation due to differences in computational mechanisms, such as quantum bits being inherently probabilistic distributions.
We use this distribution for LLP’s class-ratio supervision, providing a seamless learning framework without requiring any reinterpretation of what probabilistic qubits correspond to in conventional methods.
We evaluate Q-LLP on standard image benchmarks such as CIFAR-10, STL-10, and SVHN under weak supervision based on class proportions.
Q-LLP achieves competitive or superior accuracy, whereas conventional LLP baselines generally decline in generalization performance in small datasets.
Our results show that Q-LLP takes theoretical advantage of quantum algorithms by reducing the information loss introduced through quantum translation.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 8478