Self-Supervised Quantum Representation Learning on a Trapped-Ion Quantum Computer

Access to labeled datasets remains a key bottleneck for training and scaling machine-learning models. In this work, we demonstrate contrastive self-supervised pretraining of quantum representations executed \textit{in situ} on a programmable trapped-ion quantum computer. Classical images are encoded as quantum states using a trainable data encoder and variational circuits. During pretraining, the model learns invariances from unlabeled examples by comparing image pairs directly on hardware; similarity for the contrastive objective is obtained from measured quantum state overlaps, rather than from a classically projected feature space. Next, we fine-tune the pretrained representation for supervised classification of image families using limited labeled data. Relative to identical circuits trained from random initialization, contrastive pretraining yields higher mean test accuracy and substantially reduced run-to-run variability, with the strongest gains in low-label regimes. The learned invariances also transfer beyond the specific unlabeled samples used for pretraining, supporting generalization to previously unseen perturbations.