Encoder Architecture for Label-Efficient Quantum Representation Learning

In this work, we perform binary image classification on a trapped-ion quantum computer using a pretrained quantum encoder. The circuit consists of tunable single-qubit rotations and fixed-angle XX entanglers. We pretrain an encoder in a self-supervised manner by treating augmented versions of the same example as positive pairs, without using labels. Similarity for the contrastive objective is computed from measured quantum state overlaps rather than from a classical feature projection. The pretrained encoder improves downstream training with limited labels, increasing average accuracy and reducing run-to-run variability, and the learned invariances transfer to unseen perturbations. We discuss practical encoding strategies that map data to quantum states while keeping circuit depth compatible with hardware constraints.