Training Parameterized Quantum Circuits with Triplet Loss

Abstract

Training parameterized quantum circuits (PQCs) is a growing research area that has received a boost from the emergence of new hybrid quantum classical algorithms and Quantum Machine Learning (QML) to leverage the power of today’s quantum computers. However, a universal pipeline that guarantees good learning behavior has not yet been found, due to several challenges. These include in particular the low number of qubits and their susceptibility to noise but also the vanishing of gradients during training. In this work, we apply and evaluate Triplet Loss in a QML training pipeline utilizing a PQC for the first time. We perform extensive experiments for the Triplet Loss based setup and training on two common datasets, the MNIST and moon dataset. Without significant fine-tuning of training parameters and circuit layout, our proposed approach achieves competitive results to a regular training. Additionally, the variance and the absolute values of gradients are significantly better compared to training a PQC without Triplet Loss. The usage of metric learning proves to be suitable for QML and its high dimensional space as it is not as restrictive as learning on hard labels. Our results indicate that metric learning provides benefits to mitigate the so-called barren plateaus.