arXiv:2610.03220
Evolving Hybrid Quantum-Classical Architectures for Image Classification
Devroop Kar, Daniel Krutz, Travis Desell
Abstract
Hybrid quantum classical neural networks integrate parameterized quantum circuits (PQCs) with established deep learning architectures, but their performance depends strongly on the choice of quantum circuit architecture, a choice that remains largely manual. Most existing approaches rely on hand-designed or fixed circuit ansätze, requiring circuit structure, gate composition, and qubit connectivity to be specified in advance with no guarantee that they suit the task. This limitation is especially acute in image classification, where quantum circuits must transform features extracted by classical networks while remaining compact enough for practical training, requirements that generic, task-agnostic ansätze are unlikely to satisfy simultaneously. We extend EXAQC, an evolutionary framework for automated quantum circuit discovery, to image classification. EXAQC evolves PQCs as intermediate processing modules while retaining classical feature-extraction and prediction layers. On MNIST, Fashion-MNIST, and CIFAR-10, EXAQC achieves 98.42%, 90.62%, and 85.47% accuracy, respectively, while using comparable gate counts to other quantum architecture-search methods. Against classical networks, evolved hybrid models maintain comparable accuracy with substantially fewer trainable parameters, reaching 85.68% on CIFAR-10 with over 25$\times$ fewer parameters than a 10-layer CNN. Encoding choice also matters: rotation-based encodings (RX, RY, U3) outperform amplitude encoding by 22-25 points on CIFAR-10. These results demonstrate that automated circuit discovery yields compact quantum modules that can replace larger classical components in vision architectures while retaining competitive accuracy.
AI Summary
The paper asks whether evolutionary quantum architecture search can discover compact parameterized quantum circuits (PQCs) for image classification within hybrid models. It extends EXAQC to pipelines comprising a trainable classical feature encoder, an evolved PQC, and a classical decoder. Mutation and crossover alter circuit structure, while gradient-based training jointly optimizes classical and quantum parameters; validation performance guides selection. Simulation experiments cover full ten-class MNIST, Fashion-MNIST, and CIFAR-10, with ten independent runs, search budgets of 800–1,000 candidate circuits, and 5–10 qubits. The main comparison reports test accuracies of $98.42\pm1.03\%$, $90.62\pm0.80\%$, and $85.47\pm1.37\%$, respectively, using 20–30 gates. A separate parameter-efficiency comparison reports $85.68\%$ CIFAR-10 accuracy with 100,173 trainable parameters versus $85.50\%$ with 2,569,962 parameters for a ten-layer CNN. Encoding experiments find CIFAR-10 accuracies of $83.75$–$86.68\%$ for rotation-based encodings versus $61.38\%$ for amplitude encoding. These comparisons fix qubit count, so U3 receives three times as many encoder features as RX or RY. An appendix derives parameter-count conditions for replacing an affine layer with a hybrid module; these do not establish accuracy guarantees. The stated limitations include non-controlled comparisons with prior studies, dependence on classical feature quality, simulation-based evaluation without realistic-noise or hardware validation, and increasing evolutionary-search cost. Proposed extensions include noise-aware objectives, more efficient distributed search, and multi-objective optimization of accuracy and circuit resources.
It has also been used to polish the writing, identify and correct grammatical errors and typos in the text.
Central claim may not hold: The paper explicitly selects representative circuits using test performance and reports best-run parameter-efficiency accuracies without variability, weakening the independence and reliability of those comparisons. The parameter-efficiency experiments lack a comparably compressed and comparably searched classical replacement, so they establish comparisons with selected baselines rather than a benefit attributable to the evolved quantum module. The encoding comparison changes encoder output dimensionality and projection parameters between U3 and RX/RY, confounding encoding choice with the classical representation. For the described U3 input encoding on |0>, the lambda feature has no effect on the state or subsequent outputs, contradicting the claimed three-feature-per-qubit encoding.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- Under Review at The Fifteenth International Conference on Learning Representations 2027