EXAQC extends evolutionary quantum circuit discovery to image classification, reaching 85.68% on CIFAR-10 with over 25x fewer parameters

EXAQC is extended to image classification, where it evolves parameterized quantum circuits automatically and uses them as intermediate processing modules between classical feature-extraction and prediction layers. Most existing approaches rely on hand-designed or fixed circuit ansatze that require circuit structure, gate composition, and qubit connectivity to be specified in advance with no guarantee they suit the task; this work moves circuit-structure choice from manual specification to automated search. The abstract states that EXAQC is an existing evolutionary framework for automated quantum circuit discovery, extended here to image classification, with experiments on MNIST, Fashion-MNIST, and CIFAR-10.

The evolved hybrid models achieve 98.42%, 90.62%, and 85.47% accuracy on the three image datasets while using gate counts comparable to other quantum architecture-search methods. It reports concrete accuracies while keeping gate counts comparable to other quantum architecture-search methods, indicating that automatically discovered compact quantum modules are usable for image classification. The abstract reports specific accuracy values for the three datasets and uses gate count as the comparability metric against other quantum architecture-search methods.

Against classical networks, the evolved hybrid models maintain comparable accuracy with substantially fewer trainable parameters, reaching 85.68% on CIFAR-10 with over 25x fewer parameters. It positions quantum modules as compact components that can replace larger classical components, rather than only competing with classical models on accuracy. The abstract gives a parameter comparison against a 10-layer CNN (over 25x fewer) together with the 85.68% CIFAR-10 accuracy.

Encoding choice matters: rotation-based encodings (RX, RY, U3) outperform amplitude encoding by 22-25 points on CIFAR-10. It identifies data-encoding choice as a key design variable in hybrid quantum-classical image classification, beyond circuit structure itself. The abstract reports a 22-25 point gap on CIFAR-10 between rotation-based encodings and amplitude encoding.

The work targets image classification and is meant for researchers and practitioners who want to replace manual quantum circuit design with automated search and to insert compact quantum modules between classical feature-extraction and prediction layers. Its conclusions rest on three benchmark datasets, MNIST, Fashion-MNIST, and CIFAR-10, and use gate count and trainable parameters as the comparison dimensions against other quantum architecture-search methods and classical networks.

The available text is at the abstract level and does not include details of the evolutionary algorithm, circuit representation, search cost, training configuration, or statistical significance, so those specifics remain open questions. The 22-25 point advantage of rotation-based encodings over amplitude encoding is reported on CIFAR-10, and how it holds on other datasets or with other encoding schemes is a further question. The parameter-efficiency comparison uses a 10-layer CNN as reference, and relative behavior under other classical architectures is likewise worth watching.