acceptodds
Under review as a conference paper at ICLR 2027

Evolving Hybrid Quantum–Classical Architectures for Image Classification

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 ans\"atze, 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 ans\"atze 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 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.

Primary area:
optimization
Keywords:
neuroevolution, genetic algorithm, quantum machine learning, image classification

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_QasiOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper extends the EXAQC evolutionary framework to automate quantum circuit discovery within hybrid quantum-classical architectures for image classification tasks. It processes high-dimensional images by combining classical encoders, evolved parameterized quantum circuits, and classical decoders. Evaluated across MNIST, Fashion-MNIST, and CIFAR-10, the methodology demonstrates that automatically discovered quantum modules maintain competitive accuracy while using significantly fewer trainable parameters than pure classical baselines.

Critical strengths
  1. Automated discovery of PQC structure avoids hand-crafted ansatz bottlenecks, enabling hybrid networks to drastically compress trainable parameter count on CIFAR-10 compared with traditional CNNs without severe loss in accuracy. 2. Jointly optimizing classical feature extraction with rotation-based U3 encoding allows quantum layers to operate efficiently on compressed low-dimensional features, which is essential for scaling QML to complex computer vision tasks.
Critical weaknesses
  1. It is not fully clear which specific features of multi-channel spatial images make existing QAS frameworks designed for low-dimensional data unsuitable, as classical encoders handle all spatial compression before passing vectors to the quantum module. 2. The experimental comparisons in Table 1 mix drastically different dataset setups, numbers of classes, and downscaling protocols, which obscures direct quantitative baseline comparisons.
Questions for authors
  1. Can the methodology provide a top-level pseudocode/algorithm listing the exact steady-state evolutionary loop, specifying how mutation, crossover, and Lamarckian parameter inheritance operate during candidate selection in Section 2.3? 2. In Section 1 and Section 2.1, what specific domain features of image classification make prior QAS architectures designed for general low-dimensional datasets inefficient, given that classical CNN encoders compress raw pixels into feature vectors prior to quantum encoding? 3. Regarding Section 4.1 and Table 1, could the paper clarify how fairness is preserved when comparing EXAQC on 10-class benchmarks against prior works evaluated on 2-class or downscaled variants? 4. In Section 4.2 and Table 2, please provide a more detailed line-by-line discussion with bulleted observation points analyzing why U3 encoding scales better for complex data like CIFAR-10 despite requiring higher gate count and parameter complexity.
2 comments
AuthorsOpenReview ↗

Response to Reviewer

We thank the reviewer for their comments and address the questions raised.

  1. Steady-state evolutionary loop. We have added a top-level algorithm to Section 2.3. It will show when parents are selected, how offspring inherit trained parameters, and how each completed evaluation updates the population.

  2. Role of image structure after CNN compression. The CNN learns spatial patterns and relationships among input channels before the quantum circuit receives a feature vector. We agree that our experiments do not establish that these image properties make prior QAS methods inefficient. The question we study is how effectively circuit search can use a compact feature vector when the encoder, quantum circuit, and decoder are trained jointly. The qubit budget and encoding determine how many encoder outputs can enter the circuit and how they are represented. We will revise Sections 1 and 2.1 to make that distinction explicit.

  3. Comparability of the results in Table 1. We agree that results on two-class or downscaled versions are not directly comparable with the full ten-class tasks. Table 1 was meant to show that EXAQC can solve the full ten-class problems that most earlier methods reduce to two classes, not to claim higher accuracy than them. It also represents a significant advancement over prior state of the art works -- EXAQC can solve the full problems while prior methods could not. In the revision, we split Table 1 into reduced-task and ten-class groups and compare accuracy within the two groups, where EXAQC is best on Fashion-MNIST and CIFAR-10 and close on MNIST. We have revised Section 4.1 accordingly. Most prior quantum architecture search methods reduce these datasets to two classes and downscaled images, because the field is still at an early stage. We do not expect current noisy intermediate-scale quantum machines to beat state of the art classical methods, and the question of advantage will need fault-tolerant, error corrected quantum computers. Before advantage can be studied on that hardware, the community needs search methods that work on the full problems.

  4. Encoding results and the CIFAR-10 trade-off. U3 accepts three encoder outputs per encoded qubit, compared with one for RX or RY. This wider interface gives the encoder more ways to map learned features into quantum states, while the larger evolved circuit may contribute additional capacity. We have added a section to explain this within the given space constraints.

I would like to thank the authors for the responses. In light of the rebuttal and other reviews, I keep my score.

Reviewer_VGNMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors take EXAQC, their own evolutionary framework for quantum circuit architecture search, and point it at image classification: a CNN encoder compresses an image into a dq-dimensional feature vector, an evolved parameterized quantum circuit transforms it, and a linear decoder produces the class logits. They report 98.42 / 90.62 / 85.47 on the full ten-class MNIST, Fashion-MNIST and CIFAR-10, then compare four input encodings (amplitude, RX, RY, U3) and argue that the hybrid models match classical baselines with far fewer trainable parameters.

Critical strengths

1Putting evolutionary search on the full ten-class image tasks rather than two-class subsets is the right instinct, and it makes the question of handling high-dimensional inputs meaningful instead of decorative. 2The coverage is reasonably complete: one section each for cross-dataset comparison, encoding ablation, and parameter efficiency, so the experimental skeleton is easy to follow. 3Section 4.2 ends by acknowledging that fixing the qubit count means U3 receives three times as many input features, and states the trade-off explicitly. That kind of self-awareness is rarer in this literature than it should be. 4Figure 3 shows an evolved circuit that is genuinely not a template ansatz: distant Ising couplings in an irregular arrangement. It is a nice sanity check that the search is doing something structural.

Critical weaknesses

1The contribution is a port. Section 2 is a restatement of EXAQC, and the paper itself writes that it "extends EXAQC (Kar et al., 2026b)". What is new is a CNN encoder in front, a U3 encoding option, and three datasets. That is thin for this venue. 2.The comparison in Table 1 does not hold. EXAQC is evaluated on ten classes; every architecture-search baseline in the table is marked Cls. = 2. The caption concedes that accuracies are "not necessarily directly comparable", and yet the abstract, RQ1 and Section 5 all claim that EXAQC "exceeds the compared architecture-search methods on CIFAR-10". Both statements cannot stand. A genuine comparison needs matched class counts, matched preprocessing and a common resource-counting convention. 3.The headline numbers contradict one another, in three places. Any one of these is enough to undermine the reported best results: CIFAR-10 has three different "best" values: 85.47 in Table 1 at 30 gates / 21 parameters, 86.68 in Table 2 for U3 at 60 gates / 40 parameters, and 85.68 in Table 3. Section 4.3 states MNIST accuracy of 96.13%, while Table 3 and Section 5 both say 96.83%. Section 4.2 says CIFAR-10 evolved circuits contain 42-60 gates, but Table 1 records 30 gates / 21 parameters for CIFAR-10, identical to Fashion-MNIST. Which run does that circuit come from? The Table 3 caption says "using RY angle encoding", but Table 2 gives RY on CIFAR-10 as 84.37. 4.Model selection used the test set. Section 3 states: "Representative evolved circuits are selected using test performance." If 85.47 or 85.68 was picked out of ten runs by test accuracy, what is reported is not generalisation. The selection protocol needs to be spelled out, and the headline number should come from selection on validation only. 5.The one control that matters is missing. Keep the same encoder and decoder and the same parameter budget, and replace the PQC with a linear bottleneck layer. What accuracy does that reach? Without this number there is no evidence that the quantum module contributes anything, and the current "EXAQC + MLP vs MLP" rows differ by more than one layer. The Fashion-MNIST U3 row makes the point sharply: four gates and four parameters yielding 87.26% is hard to read as a circuit doing substantive work. It looks like the encoder and decoder are doing the work.

Questions for authors

1.With encoder and decoder held fixed and the PQC replaced by a linear bottleneck of comparable parameter count, what accuracy is reached on CIFAR-10 and Fashion-MNIST? 2.What is the relationship between the three CIFAR-10 numbers in Tables 1, 2 and 3? Which is the primary result, and was any selection done on the test set? 3.Can you add a dimension-matched amplitude-vs-angle comparison at equal dq and equal qubit count? 4.What are the four U3 gates on Fashion-MNIST? Does the gate count include encoding gates and the single-qubit rotations before measurement? 5.Will Appendices A and B be included with the revised version? 6.What is the total GPU-hour cost of the evolutionary search, and how does it compare with a single training run of the baselines? 7.Is the gap between 96.83 and 97.48 on MNIST significant given a standard deviation of about 1.03?

2 comments
AuthorsOpenReview ↗

Response to Reviewer

We thank the reviewer for their comments. We address the weaknesses and questions cited by the reviewer.

Weakness

  1. The contribution is not simply a port of EXAQC, before this work, EXAQC only utilized basic angle, amplitude and basis encoding. This work adds neural encoders and decoders into EXAQC’s evolutionary loop. For each candidate circuit, EXAQC connects the encoder to the circuit inputs and the circuit measurements to the decoder, then trains all three components jointly. We also extend Lamarckian inheritance to the classical networks. Mutation copies their trained weights, crossover recombines them, and offspring pass on the weights they learn during training. This allows the feature extractor and readout to adapt as the circuit architecture evolves.

  2. We thank the reviewer for pointing this out. Prior QAS methods train on small subsets (e.g., 128–2,000 images) and heavily downscaled inputs (2×2 to 8×8 pixels); QNAS evaluates 72 candidates for two epochs on a data subset. EXAQC trains every candidate on the full training set which is a significantly more challenging problem. Please note that to our knowledge this work is the first to present full size, full class results on these datasets for quantum architecture search, and only a single other architecture search work presents full size, full class results for MNIST and CIFAR-10, which our method outperforms. We have added 2-class results as well fot better comparison.

  3. We thank the reviewer for catching these inconsistencies, which come from configurations and statistics not being stated clearly. Tables 1 and 2 report the mean over 10 runs, while Table 3 reports the best test result over 10 runs with RY encoding. The encoders also differ: for MNIST and Fashion-MNIST, Table 1 uses a one-layer CNN encoder and Table 2 an MLP encoder, while Table 3 uses an MLP for MNIST and a CNN for Fashion-MNIST, as labelled. The three CIFAR-10 values therefore measure different things: 85.47 is the Table 1 mean with RY encoding with a specific encoder, 86.68 is the U3 mean in Table 2, and 85.68 is the best RY run in Table 3, consistent with the RY mean of 84.37. We now state the encoder, encoding and statistic in every caption.

  4. We understand the confusion. The model selection was based on validation performance, but the final reported metric is on the test partition.

Answers to the Author Questions

  1. Matched linear bottleneck control. Our existing comparisons evaluate complete hybrid and classical architectures. The 2 MLP Layers were replaced by the PQC; Table 3 highlight that very comparison. We have also added a bottleneck control that preserves the encoder and decoder architectures and replaces the PQC with a classical layer of the same width. On Fashion-MNIST, the classical bottleneck achieves 80.83% accuracy with 6,896 trainable parameters, compared with 83.18% and 5,362 parameters for the six-qubit hybrid. The hybrid therefore gains 2.35 percentage points while using 22.24% fewer parameters. This provides evidence that the quantum module contributes beyond classical compression in this configuration.

  2. Relationship between the CIFAR-10 results and model selection. Table 2 reports the encoding study, in which U3 achieves 86.68%86.68\% and RY achieves 84.37%84.37\% mean test accuracy. The 85.47%85.47\% in Table 1 represents mean testing accuracy across runs on a different configuration and 85.68%85.68\% in Table 3 points to the best result obtained using a 3-layered CNN encoder.

  3. Dimension-matched amplitude-versus-angle comparison. We added a comparison at equal encoder output dimension dq=8d_q=8 for amplitude and RX/RY encoding. On CIFAR-10, amplitude achieves 83.17±0.0183.17 \pm 0.01%, compared with 83.67±0.4683.67 \pm 0.46% for RX and 83.84±0.5883.84 \pm 0.58% for RY. This complements the existing equal-qubit comparison and shows that amplitude encoding becomes competitive under the dimension-controlled setting. The two controls are evaluated separately, with three qubits for amplitude and eight for RX/RY in the equal-dimension experiment.

  4. The four Fashion-MNIST gates and gate accounting. U3 denotes the input-encoding configuration; it does not imply that the four evolved operations are themselves U3 gates.

  5. Appendices A and B. Appendices A and B were included with the supplementary material.

  6. GPU-hour cost of the evolutionary search. Each of the MNIST and Fashion-MNIST runs took around 43 and 49 hours respectively to run while CIFAR-10 took 72 hours to complete on a single GPU. Classical baselines ran within 2 hours.

  7. Significance of the MNIST accuracy gap. The point is that with over 75% less parameters the hybrid quantum circuit can reach similar performance to classical baselines.

Reviewer_VGNMOpenReview ↗

Thank you

I apologize for the delayed response. Thank you very much for the additional materials and clarifications. I have read them carefully, and they are helpful and do improve the manuscript. However, the core novelty and practical utility ultimately determine the upper bound of the paper’s contribution. After careful consideration, I will keep my current score.

Reviewer_VGg2OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper extends previous evolutionary approaches for quantum circuit design in hybrid quantum machine learning to image classification tasks. The authors report empirical results across MNIST, Fashion MNIST, and CIFAR-10, where the test accuracy of their proposed approach exceeds that of comparable hybrid quantum approaches and performs comparably to pure classical baselines.

Critical strengths

While the paper is generally well-written and easy to follow and focuses on the important aspect of scaling hybrid quantum machine learning approaches to larger-scale applications, it unfortunately does not provide substantially new insights or findings regarding the use of evolutionary optimization for hybrid quantum machine learning.

Critical weaknesses

The paper's central contribution appears to be extending the evolutionary circuit design approach introduced by Kar et al. (2026) to image classification tasks. Given that the approach builds on a well-studied hybrid setup with classical feature extraction and post-processing (e.g., Mari et al., 2019), the evaluation is limited to the number of parameters and test accuracy on common image classification benchmarks, and the results are apparently obtained from noise-free simulation, with no consideration of current or near-term noisy hardware, the paper does not provide any substantial new insights.

In addition, it lacks a comparison with, and potential integration of, current state-of-the-art classical approaches to image classification, such as Vision Transformers. Overall, despite the limited comparison to basic CNNs and MLPs, the proposed approach offers little performance improvement and, in some cases, even worsens performance, despite the added complexity of a hybrid setup with meta-optimization.

Finally, regarding the comparison of the number of parameters, I am unconvinced that a reduction in parameters is an appropriate argument for improvement here, especially given that the type, complexity, and realizability of parameters (classical vs. quantum) differ vastly and should not be quantitatively compared.

Questions for authors

How would the authors assess the computational overhead of the proposed approach, which combines meta-optimization with hybrid quantum-classical execution, relative to the seemingly deliberately limited classical baselines used for comparison?

2 comments
AuthorsOpenReview ↗

Response to Reviewer

We thank the reviewer for raising these points.

What the study adds. We first would like to highlight that the purpose of this paper was not to compete against purely classical models, but rather to demonstrate what we feel is a significant advancement in quantum machine leaning, and especially quantum architecture search, which highlights our methods future potential for use in hybrid systems, especially as hardware advances and improves. To our knowledge, this work presets results for the first quantum architecture search method that is capable of and reports results on full sized, the full class MNIST, CIFAR-10 and FashionMNIST datasets. We also improve on results from the only other work reporting results on the full sized problem, which used a fixed architecture (HQNN-Parallel) with fewer quantum gate parameters. We feel like this does add significance to our work, highlight its potential for future systems, and also showing that quantum components in hybrid models can potentially provide a significant reduction in trainable parameter complexity.

We would like to expand on this point. There is no reason to believe that current noisy intermediate-scale quantum (NISQ) machines, or the tradecraft the community has developed so far for quantum machine learning, can beat state of the art classical methods such as Vision Transformers, and we do not claim that our approach does. This agrees with recent benchmarking studies, e.g., Bowles, Ahmed and Schuld (arXiv:2403.07059), where out-of-the-box classical models generally outperform the quantum models on small-scale binary classification tasks, and the best-known provable speed-up for learning from classical data, e.g., Liu, Arunachalam and Temme (Nature Physics 17, 2021), relies on fault-tolerant hardware. We view quantum machine learning as a nascent field, and works such as this one are building the literature, tradecraft and tools that help us understand how to do quantum machine learning, i.e., how to search for, train and evaluate hybrid models on full sized, multi-class problems rather than the reduced versions used in most prior work. Once fault-tolerant, error corrected quantum computers exist we can then explore advantage, and leverage the methodology developed by the community over the years to find where quantum computers add value to machine learning. We feel works like this are needed before the value of quantum computers for machine learning can be tested.

We also clarify the purpose of the comparisons, how models were selected, and what the reported parameter counts tell us about the hybrid models.

Why we chose these baselines. Table 1 places EXAQC alongside published quantum image classifiers and quantum architecture-search methods. It shows what EXAQC achieves on the full ten-class tasks with compact circuits. Table 3 addresses a separate question about the trade-off between accuracy and model size. On CIFAR-10, the hybrid model reaches 85.68% with about 100,000 trainable parameters, while the 10-layer CNN reaches 85.50% with 2.57 million showing that a hybrid quantum model can reach accuracies comparable to classical CNN architectures with 10×10 \times more parameters. For the same reason, we did not include Vision Transformers or other state of the art classical models as baselines. Since we do not claim the hybrid model is competitive with these methods, such a comparison would not answer the question we study. In the CNN and MLP baselines the quantum circuit replaces a known classical component, so that we can compare the two at a similar accuracy.

What parameter counts tell us. We agree that a trainable quantum angle and a classical weight can have very different execution costs. The total parameter count describes how many values the complete model learns, including the encoder, circuit, and decoder. Paired with test accuracy, it describes the model-size trade-off. While current quantum hardware may limit this, this presents the potential of such methods for future hardware.

Noise-free simulation. We follow a simulation based approach to compare against works which employ similar backends. As an evolutionary approach, EXAQC is well suited for multi-objective optimization which also incorporates noise. We leave complete runs on noisy hardware (simulated or real) as part of future work. We also note that noise-free simulation is the setting in which the methodology will eventually be used. On an error corrected quantum computer the logical qubits are designed to behave close to noiselessly, so a noiseless simulation is a reasonable model of the fault-tolerant regime where we expect quantum advantage to be explored; results on NISQ hardware would in large part reflect the noise of the device used.

Thank you for the detailed response and the clarifications on the scope of the work, the choice of baselines, and the use of noise-free simulation. I appreciate the authors’ candid admission that the approach is not intended to compete with state-of-the-art classical methods and that no near-term advantage is expected.

However, this framing reinforces my central concern rather than resolving it. If the contribution is not performance, it must lie in new methodological insights into quantum architecture search. The response restates the scaling to full-size datasets but does not point to such insights beyond extending existing work (Kar et al., 2026; Mari et al., 2019).

On parameter efficiency, I appreciate the acknowledgement that classical and quantum parameters differ substantially in cost, yet parameter count remains the main argument for efficiency. Pointing to future fault-tolerant hardware does not strengthen this argument, since the overhead of implementing parameterised rotations on logical qubits would make a direct quantitative comparison even less meaningful. Similarly, I agree that noiseless simulation approximates the fault-tolerant regime, but the claimed practical benefits remain speculative.

Finally, my question about the computational overhead of the meta-optimisation combined with hybrid execution, compared with the classical baselines, remains unanswered.

In summary, I appreciate the clarifications, but the central weaknesses remain, and I maintain my assessment.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.