Evolving Hybrid Quantum-Classical Architectures for Image Classification

Devroop Kar; Daniel Krutz; Travis Desell

cs.NEOct 2, 2026

Evolving Hybrid Quantum-Classical Architectures for Image Classification

Authors: Devroop Kar, Daniel Krutz, Travis Desell

Organizations: Rochester Institute of Technology Rochester, NY 14623, USA

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.

Figures & tables

Figure 1: Overview of the proposed EXAQC hybrid quantum classical architecture. A CNN encoder compresses the input image into a lower-dimensional representation that is encoded and processed by an evolved PQC. Quantum measurements are subsequently mapped to class predictions by a classical decoder.

Panel of Figure 2 — Figure 2: EXAQC evolutionary search workflow for hybrid image classification. Parent circuit genomes are selected and modified through mutation or crossover to create offspring architectures. Each offspring is instantiated within the hybrid model and its classical and quantum parameters are optimized using gradient-based training. Validation performance determines whether the offspring is inserted into the steady-state population.

		MNIST		Fashion-MNIST		CIFAR-10
Method	Cls.	Acc.	G/P	Acc.	G/P	Acc.	G/P
Hybrid DQAS a ( Afane et al., 2026 )	2	$97.97\pm 0.40$	–	$93.86\pm 1.70$	–	$71.82\pm 1.60$	–
DQAS b ( Zhang et al., 2022 )	2	$96.28\pm 3.90$	–	$93.40\pm 0.09$	–	$71.12\pm 5.70$	–
QuantumDARTS b ( Wu et al., 2023 )	2	$95.66\pm 3.30$	–	$92.20\pm 0.06$	–	$70.14\pm 1.80$	–
EQAS ( Zhang and Zhao, 2023 )	2	$94.00$	22/ 9	$75.00$	24/20	–	–
EQNAS e ( Li et al., 2023a )	2	$87.35$	43/20	$84.15$	41/18	$59.65$	43/20

Table 1: Comparison with prior quantum and quantum-architecture-search approaches on image classification: test accuracy (%) and quantum circuit resources (G/P = gates / quantum gate parameters). Cls.- specifies how many classes were actually used from the dataset. Accuracies are not necessarily directly comparable, as prior studies differ in number of classes, dataset subsets, preprocessing, and quantum input representations. “–” indicates a result or resource count that was not reported.

Dataset	Quantum Encoding	Test Acc. (%)	# Gates	# Params.
MNIST	Amplitude	95.65 $\pm$ 0.1	8	5
	RX	96.22 $\pm$ 0.1	11	6
	RY	96.27 $\pm$ 0.2	17	9
	U3	96.42 $\pm$ 0.2	14	14
Fashion-MNIST	Amplitude	82.19 $\pm$ 0.2	5	6
	RX	88.01 $\pm$ 0.8	12	6

Table 2: Comparison of quantum encoding strategies for EXAQC across the image-classification benchmarks. Results are reported over multiple independent runs as mean

\pm

standard deviation.

Dataset	Model	Test Acc. (%)	# Trainable Params.	Param. Reduction ( $\mathbf{R_{P}}\%$ )
MNIST	MLP $(784\times 512\times 128\times 10)$	97.48	468874	78.36 ( $\downarrow$ )
MNIST	EXAQC + MLP	96.83	101458	–
Fashion-MNIST	MLP $(784\times 512\times 128\times 10)$	88.03	468874	76.67 ( $\downarrow$ )
	EXAQC + MLP	86.95	109421	–
	1-layer CNN	89.47	74922	54.83 ( $\downarrow$ )
	2-layer CNN	90.87	79418	57.38 ( $\downarrow$ )

Table 3: Comparison of classical and EXAQC-based hybrid models in terms of classification performance and parameter efficiency. Parameter reduction is calculated relative to the corresponding classical baseline for each dataset. Best Results are reported over 10 independent runs using RY angle encoding.

Figure 3: Evolved EXAQC hybrid architecture for CIFAR-10. A three-block CNN encoder compresses each image to 8 features, which are RY-encoded onto 8 qubits.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 4 — Figure 4: Representative samples from the three image-classification benchmarks considered in this work. Each panel contains one example from each of the ten classes. MNIST and Fashion-MNIST consist of $28\times 28$ grayscale images, whereas CIFAR-10 consists of $32\times 32$ RGB images.

Panel of Figure 5 — Figure 5: Evolution of classification loss during EXAQC search on (a) MNIST, (b) Fashion-MNIST, and (c) CIFAR-10 across 10 independent runs. Curves show the mean loss for the top-performing individuals, the best individual, and the overall population, with shaded regions indicating $\pm$ one standard deviation across runs.

		MNIST	Fashion-MNIST	CIFAR-10
exponential crossover	global best	0.027	0.039	0.024
	inserted	0.227	0.336	0.287
	discarded	0.747	0.625	0.689
reorder gate	global best	0.016	0.073	0.041
	inserted	0.404	0.616	0.475
	discarded	0.580	0.311	0.484

Appendix

Table 4: Performance rates for each mutation and crossover operation on the MNIST, Fashion-MNIST, and CIFAR-10 image-classification tasks. Best performance (high for global-best and inserted offspring, low for discarded offspring) is shown in bold italics, second best in bold, and third best in italics.

Panel of Figure 6 — Figure 6: Evolution of quantum circuit complexity during EXAQC search on (a) MNIST, (b) Fashion-MNIST, and (c) CIFAR-10. Solid lines show the gate and parameter counts of the best-so-far circuits, while dashed lines show the average counts of generated circuits. Shaded regions indicate variation across independent runs.

Figure 8: Fashion MNIST Genome Representation

Explore similar work

Q-PhotoNAS: Hybrid Quantum Neural Architecture Search Framework on Photonic Devices
May 21, 2026Farah Elnakhal, Alberto Marchisio, Nouhaila Innan +2Hybrid Quantum-Classical ML Hardware-Aware NAS
Science Division, New York University Abu Dhabi, UAE · eBrain Lab, Division of Engineering, New York University Abu Dhabi, PO Box 129188, Abu Dhabi, UAE · Center for Quantum and Topological Systems, NYUAD Research Institute, New York University Abu Dhabi, UAE +1
Do emulated quantum circuits change what CNNs look at? Performance and explainability comparison in medical image classification
Jul 23, 2026Guillermo Rubiños Rodríguez, Martín Ottavianelli, Mateo Alonso +4Quantum Convolutional Neural Networks Quantum-Inspired ML
Instituto Tecnol´ogico de Galicia, Cant´on Grande 9, Planta 3, 15003 A Coru˜na, Spain · Centro Singular de Investigaci´on en Tecnolox´ıas Intelixentes (CiTIUS), Universidade de Santiago de Compostela, Santiago de Compostela, 15782, Spain · Department of Neuroradiology, LMU University Hospital, Ludwig Maximilian University of Munich, Marchioninistraße 15, 81377 Munich, Germany +1
Hybrid Classical-Quantum Transfer Learning with Noisy Quantum Circuits
Mar 17, 2026D. Martín-Pérez, F. Rodríguez-Díaz, D. Gutiérrez-Avilés +2Quantum Machine Learning Hybrid Quantum-Classical ML
Data Science and Big Data Lab, Pablo de Olavide University, Seville, Spain · Department of Computer Science, University of Seville, Seville, Spain

May 21, 2026J/K move · Enter open · S save

Q-PhotoNAS: Hybrid Quantum Neural Architecture Search Framework on Photonic Devices

Farah Elnakhal, Alberto Marchisio, Nouhaila Innan, Gabriel Falcao+1

Science Division, New York University Abu Dhabi, UAE · eBrain Lab, Division of Engineering, New York University Abu Dhabi, PO Box 129188, Abu Dhabi, UAE · Center for Quantum and Topological Systems, NYUAD Research Institute, New York University Abu Dhabi, UAE+1

Hybrid Quantum-Classical ML Hardware-Aware NAS Photonic Computing

Open paper AskPDF arXiv

Photonic quantum computing is a promising platform for scalable quantum machine learning, but designing effective hybrid architectures remains challenging under hardware and optimization constraints. Existing approaches rely on manually tuned architectures that fail to account for the collaboration between classical preprocessing, phase encoding, and photonic circuit structure, limiting both accuracy and hardware compatibility. In this paper, we propose a neural architecture search framework for hybrid photonic quantum-classical models that combines genetic algorithm-based search with learnable quantum phase encoding to systematically explore the joint design space of classical and quantum components. Our framework encodes 19 hyperparameters across six gene groups and evolves a population of hybrid architectures using group-based crossover, per-gene mutation, and elitism, evaluating each candidate on a short training budget before full retraining of the best found design. We evaluate our framework on two image classification benchmarks, Digits and MNIST, achieving final validation accuracies of 99.44% and 98.78%, respectively, with first-principles execution time estimates on the Quandela Ascella photonic QPU projecting single-image inference at 67 ms (Digits) and 149 ms (MNIST). Our quantum contribution analysis further shows that the photonic layer extracts non-redundant features orthogonal to the classical pathway, providing a measurable accuracy advantage over classical-only baselines. Our results demonstrate that automated architecture search is both practical and impactful for hybrid photonic systems, opening the way for systematic design space exploration of quantum AI on photonic devices.

Abstract

Explore similar work

Q-PhotoNAS: Hybrid Quantum Neural Architecture Search Framework on Photonic Devices

Do emulated quantum circuits change what CNNs look at? Performance and explainability comparison in medical image classification

Hybrid Classical-Quantum Transfer Learning with Noisy Quantum Circuits