⚛️ quantum physics

Evolving Hybrid Quantum-Classical Architectures for Image Classification

This paper extends the EXAQC evolutionary framework to automate the discovery of hybrid quantum-classical architectures for image classification, demonstrating that evolved parameterized quantum circuits can achieve competitive accuracy on MNIST, Fashion-MNIST, and CIFAR-10 while significantly reducing the number of trainable parameters compared to classical deep learning models.

Original authors: Devroop Kar, Daniel Krutz, Travis Desell

Published 2026-10-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Devroop Kar, Daniel Krutz, Travis Desell

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the quiet race to build smarter computers, a new frontier has emerged where the rigid logic of silicon chips meets the fluid, probabilistic nature of quantum mechanics. For years, scientists have been trying to teach machines to recognize patterns, from spotting a cat in a photograph to identifying a tumor in an X-ray. Traditional computers do this by processing vast amounts of data through layers of mathematical filters, a method that works well but requires immense energy and memory. Quantum computers, which use the strange rules of physics that govern atoms, promise to solve these problems with far less effort, but they are notoriously difficult to control. The biggest hurdle is that the "circuits" inside these quantum machines—the specific arrangements of gates that process information—must be designed with extreme precision. If the design is even slightly off, the computer fails to learn, or worse, stops learning entirely. For a long time, researchers had to hand-craft these designs, guessing which arrangement of quantum parts would work best for a specific task, a process that was slow, expensive, and often ineffective.

A team of researchers at the Rochester Institute of Technology has taken a different approach, one that lets the computer design itself. They developed a system called EXAQC, which uses an evolutionary process similar to natural selection to discover the best possible quantum circuit for a given job. Instead of a human engineer drawing a blueprint, the system generates thousands of random circuit designs, tests them, and keeps the ones that perform best. It then mixes and mutates these successful designs to create even better versions, repeating the cycle until it finds a solution that is both accurate and efficient. The researchers tested this system on the difficult task of image classification, asking the machine to look at pictures of handwritten numbers, clothing items, and everyday objects like cars and animals, and sort them into the correct categories.

The results of this experiment reveal a striking shift in how we might build future intelligent systems. When the researchers applied their evolutionary system to the task of recognizing handwritten digits, it achieved a success rate of nearly 98.5 percent. When the task became harder, moving to images of clothing, the system still managed to get about 90.6 percent correct. The true test came with the most complex dataset, a collection of colorful images of ten different object classes. Here, the system reached an accuracy of 85.5 percent. What makes these numbers particularly significant is not just that the system worked, but how it worked. The researchers found that the quantum circuits the system evolved were incredibly compact, using only a tiny fraction of the resources required by traditional computer models. In fact, on the most complex dataset, the hybrid system achieved the same level of accuracy as a massive, ten-layer deep neural network while using over ninety-six percent fewer adjustable settings, or parameters, to do the job.

The study also uncovered that the way data is fed into the quantum machine matters more than previously thought. The researchers tried different methods of translating a standard digital image into a quantum state. They found that for simple images, the method didn't make much of a difference. However, for the complex, colorful images, the choice of translation method was critical. One specific method, which uses rotation-based encoding, allowed the system to understand the images much better than the traditional amplitude-based method, improving accuracy by more than twenty percentage points. This suggests that as the tasks become harder, the way we prepare the data for the quantum processor becomes just as important as the processor itself.

Perhaps the most surprising discovery was how the system handled the sheer volume of information in the images. A raw image from the complex dataset contains over three thousand data points, far too many for current quantum hardware to handle directly. The researchers solved this by using a standard computer component to compress the image into a small, manageable summary before passing it to the quantum part. The evolutionary system then learned how to process this summary. It turned out that the quantum circuits evolved to handle these compressed summaries were not simple, repetitive structures. Instead, they developed complex, irregular patterns of connections that looked nothing like the fixed designs humans usually create. These evolved circuits were able to extract the necessary information from the compressed data with remarkable efficiency, proving that an automated search can find solutions that human designers might overlook.

The researchers were careful to note that these results come from computer simulations, not from running the circuits on actual quantum hardware, which is still in its early stages and prone to errors. However, the consistency of the results across ten separate runs of the experiment suggests that the findings are robust and not just a lucky fluke. The system did not just find a solution; it found a solution that was significantly more efficient than existing methods. By combining a classical computer's ability to extract features with a quantum computer's ability to process them, the researchers demonstrated a path forward that avoids the need for massive, energy-hungry models. The work shows that we do not need to wait for perfect quantum hardware to see benefits; by evolving the right architecture for the specific task at hand, we can build hybrid systems that are powerful, compact, and ready for the challenges of real-world data.

Technical Summary: Evolving Hybrid Quantum–Classical Architectures for Image Classification

Problem Statement

Hybrid quantum-classical neural networks (HQNNs) integrate parameterized quantum circuits (PQCs) with deep learning architectures to process classical data. However, the performance of these models is heavily dependent on the manual selection of the quantum circuit architecture (ansatz), including gate composition, circuit depth, qubit connectivity, and data encoding. Existing approaches typically rely on fixed or hand-designed circuits that are not guaranteed to suit specific tasks. This limitation is particularly acute in image classification, where quantum circuits must transform features extracted by classical networks while remaining compact enough for practical training on near-term hardware. Generic, task-agnostic ansätze often fail to meet these simultaneous requirements, leading to issues such as barren plateaus (exponentially vanishing gradients) and poor trainability. Consequently, there is a need for an automated method to discover quantum circuit structures that are both effective for complex image data and computationally efficient.

Methodology

The authors extend EXAQC, an evolutionary framework previously used for low-dimensional benchmarks, to the domain of image classification. The proposed methodology employs a hybrid architecture consisting of three main components:

Classical Encoder ( $E_\psi$ ): A feature extractor (identity block, MLP, or CNN) that compresses high-dimensional raw images ( $C \times H \times W$ ) into a compact representation ( $d_q$ ) suitable for quantum processing.
Evolved Parameterized Quantum Circuit ( $Q_{\theta, A}$ ): A PQC where the architecture $A$ (gate composition, depth, connectivity) is discovered via evolutionary search, while the continuous parameters $\theta$ are optimized via gradient descent.
Classical Decoder ( $D_\phi$ ): A network that maps quantum measurement outcomes to class predictions.

Evolutionary Search Process:
EXAQC utilizes a steady-state evolutionary algorithm to optimize the circuit structure. Each candidate circuit is represented as a variable-length genome containing an ordered collection of gates. The search employs:

Structural Mutation: Adding, enabling, disabling, or reordering gates, and modifying qubit connectivity.
Crossover: Combining information from parent circuits using binary, exponential, or n-ary operators.
Lamarckian Inheritance: Transferring learned continuous parameters from parents to offspring to accelerate convergence.
Asynchronous Evaluation: Candidates are evaluated in parallel, with offspring immediately competing for population insertion based on validation fitness.

Quantum Encoding Strategies:
The study investigates four input encoding mechanisms:

Amplitude Encoding: Maps $d_q$ features to $n_q = \lceil \log_2(d_q) \rceil$ qubits.
Rotation Encodings (RX, RY): Maps one feature per qubit ( $n_q = d_q$ ).
U3 Encoding: Maps three features per qubit using the $U3(\theta, \phi, \lambda)$ gate, allowing $n_q = d_q/3$ .

The training objective minimizes cross-entropy loss across the joint parameters of the encoder, quantum circuit, and decoder.

Key Contributions

Extension of EXAQC to Image Classification: The paper successfully scales the EXAQC framework from low-dimensional problems to complex image datasets (MNIST, Fashion-MNIST, CIFAR-10), demonstrating that evolved circuits can find useful transformations in compressed image features.
Automated Discovery of Compact Quantum Modules: Unlike methods relying on fixed ansätze, the proposed approach evolves circuits tailored to the learned features, resulting in compact quantum modules (20–60 gates) that replace larger classical components.
Novel Use of U3 Encoding: The work introduces the use of U3 gates as a quantum encoding mechanism within an evolutionary search framework, distinguishing it from previous works that did not employ this specific encoding strategy.
Parameter Efficiency: The study provides empirical evidence that evolved hybrid models can achieve competitive accuracy with significantly fewer trainable parameters compared to purely classical deep learning baselines.

Experimental Results

The framework was evaluated on three datasets with varying complexity:

Predictive Performance:
- MNIST: 98.42% accuracy.
- Fashion-MNIST: 90.62% accuracy.
- CIFAR-10: 85.47% accuracy.
- These results exceed or match prior quantum architecture search (QAS) methods (e.g., EQNAS, AQEA-QAS) on CIFAR-10 while utilizing fewer gates.
Impact of Encoding:
- On MNIST, encoding choice had minimal impact (differences < 1%).
- On CIFAR-10, rotation-based encodings (RX, RY, U3) significantly outperformed amplitude encoding (83.75–86.68% vs. 61.38%). U3 encoding yielded the highest accuracy (86.68%) but required the largest circuits (60 gates, 40 parameters).
Parameter Efficiency:
- CIFAR-10: The EXAQC hybrid model achieved 85.68% accuracy with 100,173 parameters, matching a 10-layer CNN (85.50%) that required 2,569,962 parameters (a 96.1% reduction). It also remained within 1.4% of ResNet18 accuracy with 99.1% fewer parameters.
- Fashion-MNIST: The hybrid model exceeded a 1-layer CNN (91.11% vs. 89.47%) with 54.8% fewer parameters.
Circuit Complexity:
- Evolved circuits for CIFAR-10 were substantially larger and more entangled than those for MNIST, featuring irregular, non-layered arrangements of single-qubit rotations and multi-qubit controlled gates, differing from standard hand-designed templates.

Significance and Claims

The paper claims that automated circuit discovery via evolutionary search yields compact quantum modules capable of replacing larger classical components in vision architectures without sacrificing competitive accuracy. The significance of this work lies in:

Scalability: Demonstrating that evolved PQCs can handle the compressed feature representations of complex image datasets, bridging the gap between near-term hardware limitations and high-dimensional data requirements.
Efficiency: Establishing that hybrid models can achieve state-of-the-art performance for their class with drastically reduced parameter counts, addressing the resource constraints of current quantum hardware.
Adaptability: Highlighting that the optimal quantum architecture and encoding strategy are not universal but depend heavily on the dataset and the quality of the classical feature extraction.

The authors conclude that while the results are promising, future work must address the reliance on simulation, the need for standardized comparisons with prior art, and the joint optimization of encoders and quantum circuits. They also note that evolutionary search costs increase with circuit complexity, suggesting a need for more efficient distributed search strategies.