Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Log inSign up
Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Blog
Job Board
Q3AS

© 2026 Aqora Quantum S.A.S.

TermsPrivacyLegal Notice

arXiv:2506.23765

QMetric: Benchmarking Quantum Neural Networks Across Circuits, Features, and Training Dimensions

Silvie Illésová, Tomasz Rybotycki, Martin Beseda

Published Jun 30, 2025• 11 citations
arXivHTMLPDF

Abstract

As hybrid quantum-classical models gain traction in machine learning, there is a growing need for tools that assess their effectiveness beyond raw accuracy. We present QMetric, a Python package offering a suite of interpretable metrics to evaluate quantum circuit expressibility, feature representations, and training dynamics. QMetric quantifies key aspects such as circuit fidelity, entanglement entropy, barren plateau risk, and training stability. The package integrates with Qiskit and PyTorch, and is demonstrated via a case study on binary MNIST classification comparing classical and quantum-enhanced models. Code, plots, and a reproducible environment are available on GitLab.

DOI
10.48550/arxiv.2506.23765

Related papers

Leveraging Quantum Layers in Classical Neural Networks

Silvie Illésová

Hybrid quantum-classical neural networks represent a promising frontier in the search for improved machine learning models. This thesis explores the integration of quantum layers within classical convolutional neural network architectures, aiming to leverage quantum entanglement and feature mapping to enhance learning capabilities. A detailed methodology for constructing and training such hybrid models is presented, using PyTorch and Qiskit Machine Learning frameworks. Experiments investigate the performance impact of inserting quantum layers at different stages of the neural network pipeline. The results suggest that quantum components can introduce meaningful transformations even with a limited number of qubits, motivating further research into scalable quantum machine learning. The full implementation is made publicly available, and future work will focus on expanding experimental evaluations and publishing additional findings.

arXiv:2507.12505•Jul 16, 2025•0 citations

Benchmarking the Quantum Frontier: Evaluating QML Models on Core Learning Tasks

Priyanka Samarth, Sadhana Shashidhar, Sakshi Rajani +2 more

Advancements in quantum computing have sparked a systematic evaluation of quantum machine learning (QML) models, focusing on their behavior and performance across diverse learning tasks such as classification, regression, generative modeling, and feature compression. The primary focus is on assessing the performance and behavior of various QML architectures-including Quantum Neural Networks (QNNs), Quantum Support Vector Classifiers (QSVCs), Quantum Generative Adversarial Networks (QGANs), Quantum Convolutional Neural Networks (QCNNs), and Quantum Autoencoders (QAEs)-implemented using Qiskit's machine learning library. This evaluation is conducted using an existing benchmarking framework that ensures consistent experimental conditions and supports flexible evaluation strategies. To provide additional context and insight, these quantum models are compared against classical counterparts using widely adopted implementations. Experiments are carried out on a range of datasets, including structured and low-dimensional examples like Bars and Stripes, as well as more complex datasets like Downscaled MNIST and synthetic Hyperplanes. By providing a standardized evaluation of these models within the Qiskit framework, this study establishes a reproducible performance baseline and highlights key challenges for QML systems in the noisy intermediate-scale quantum (NISQ) era.

International Conference on AI-ML Systems (AIMLSystems)•Oct 8, 2025•0 citations

Exploring Quantum Neural Networks: A Multi-Platform Study on Quantum-Based Classification

Savita Kumari Sheoran, Vikesh Yadav, Rakesh Kumar Sheoran

Quantum-assisted models can surpass classical counterparts in specific domains by leveraging quantum superposition entanglement for more efficient data representation. This study investigates a Quantum neural network framework executed on various Quantum simulators, including ionq, Cirq, Pennylane and Quantinnum backends. The iris, wine, Indian liver patient and rain datasets are employed as benchmarking datasets to assess the efficacy of quantum-based classification models. The QNN utilises parameterized quantum circuits and quantum feature maps to learn complex patterns in data, optimizing classification decision boundaries through quantum optimization techniques. The study evaluates QNN performance across different quantum hardware platforms, assessing key metrics such as accuracy, recall, F1 score, and precision. Preliminary results indicate that QNN-based classification effectively captures feature dependencies through quantum correlations, offering competitive performance in data classification tasks. These findings contribute to the broader exploration of quantum-enhanced classification and the potential of quantum devices in real-world data analysis.

International Conference on Networks and Cryptology (NETCRYPT)•May 29, 2025•0 citations

MNIST quantum classification models implementation and benchmarking

Sirawit Butmaratthaya, Niti Buesamae, Unchalisa Taetragool

Quantum machine learning is a new field of study that aims to apply the advantages of quantum computing, such as superposition, which is one of the intriguing properties that could speed up computation, to machine learning. Like classical machine learning models, quantum machine learning models can learn from specific input data and perform specific prediction tasks. However, quantum machine learning models are currently in the form of quantum circuits with parameterizable gates. In this paper, the implementation of a quantum classification model is investigated using the two most well-known quantum machine learning development tools: Cirq with TensorFlow Quantum (TFQ) and Qiskit with PyTorch. The quantum classification circuit is trained to recognize images of handwritten digits using the MNIST dataset. However, because the quantum computer is still in development and is difficult to access and use, quantum simulators are the best alternative for the time being. The performance of the quantum classification model implemented by the two development tools is then compared in this paper experimentally using quantum simulators on two different hardware configurations. In addition, the experimental setups differ in the number of input qubits used to store the image’s classical pixel values. Finally, the wall time and CPU time used to run the training process, as well as classification accuracy, are examined. In terms of accuracy, Qiskit with PyTorch outperforms Cirq with TFQ only in the lower number of qubits but the results are similar in the higher number of qubits. The experiment results in terms of run-time are highly different. From the result, we find that Cirq is noticeably faster than Qiskit in every experiment by at least 90 times.

AIP conference proceedings•Jan 1, 2023•2 citations

Quantum Neural Networks for Pattern Recognition and Prediction

Oscar Muller, Elena Schmidt, Anna Horvath

Quantum neural networks (QNNs) -- parameterised quantum circuits trained to perform pattern recognition and predictive tasks -- represent a prominent but contested direction in quantum machine learning, offering potential parameter efficiency and expressibility benefits over classical neural networks for specific problem structures while suffering from barren plateau trainability limitations at scale. The 2024 issue of this journal established SNAD LSAD-structured ansatz (Silva et al., 2024) as the critical enabler of scalable QNN training, and QHDC (Bianchi et al., 2024) showed that QNN advantage is dataset-dependent. This paper advances the QNN field with the Quantum Neural Network Design and Evaluation (QNNDE) framework, a systematic methodology for QNN architecture design, training optimisation, and performance evaluation across four pattern recognition and prediction task classes: temporal sequence modelling (quantum recurrent circuits), spatial pattern recognition (quantum convolutional circuits), generative modelling (quantum generative adversarial networks), and time-series regression (quantum reservoir computing). QNNDE is evaluated on eight benchmark tasks spanning physics simulation, molecular property prediction, quantum state tomography, financial time-series, medical signal classification, and quantum circuit synthesis on IBM Quantum Eagle (127-qubit) and IonQ Forte (35-qubit). Quantum recurrent circuits achieve 12.4pp accuracy improvement over classical LSTMs for quantum state evolution prediction. Quantum convolutional circuits match classical CNNs on image-derived molecular patterns at 68.4% parameter reduction. Quantum GANs generate molecular conformations with 8.4% lower FID than classical GAN at equal network size. Quantum reservoir computing outperforms classical echo state networks by 18.4% on quantum-process time-series prediction.

Quantum Frontiers Journal•Mar 28, 2025•0 citations

Can the Quantum Fisher Effective Dimension Judge a Hybrid Quantum Neural Network?

Md Masum Mia, Amith Khandakar, Mohamed Arselene Ayari

Hybrid quantum neural networks (QNNs) pair a classical feature extractor with a variational quantum circuit, and the effective dimension of the circuit's quantum Fisher information matrix (QFIM) is widely used to assess such models. This scalar is often interpreted as an indicator of trainability or accuracy, although these readings are rarely tested under controlled conditions. We therefore investigate what it actually measures. We define the effective dimension as the participation ratio of the QFIM diagonal, which lies between one and the parameter count, and we evaluate it across two image datasets, four frozen backbones including a random-projection control, four circuit depths on eight qubits, and three seeds. During training, the effective dimension remains well below the parameter count, and its correlation with test accuracy ranges from -0.72 to +0.87, changing sign across seeds in half of the configurations. Increasing depth activates a larger share of the circuit directions, from 0.38 to 0.91 of the parameter count, while the initialization gradient variance remains between 1e-5 and 1e-4, so no barren plateau (BP) appears at eight qubits. At a fixed circuit, the four backbones share an initial effective dimension of 6 to 9, even though CIFAR-10 accuracy ranges from 23.14% to 86.44%. These results show that the effective dimension characterizes circuit geometry rather than model quality. We conclude with a practical protocol that assigns each diagnostic to the question it can answer.

quant-pharXiv:2610.14070•Oct 9, 2026•0 citations
As hybrid quantum-classical models gain traction in machine learning, there is a growing need for tools that assess their effectiveness beyond raw accuracy. We present QMetric, a Python package offering a suite of interpretable metrics to evaluate quantum circuit expressibility, feature representations, and training dynamics. QMetric quantifies key aspects such as circuit fidelity, entanglement entropy, barren plateau risk, and training stability. The package integrates with Qiskit and PyTorch, and is demonstrated via a case study on binary MNIST classification comparing classical and quantum-enhanced models. Code, plots, and a reproducible environment are available on GitLab.
Hybrid quantum-classical neural networks represent a promising frontier in the search for improved machine learning models. This thesis explores the integration of quantum layers within classical convolutional neural network architectures, aiming to leverage quantum entanglement and feature mapping to enhance learning capabilities. A detailed methodology for constructing and training such hybrid models is presented, using PyTorch and Qiskit Machine Learning frameworks. Experiments investigate the performance impact of inserting quantum layers at different stages of the neural network pipeline. The results suggest that quantum components can introduce meaningful transformations even with a limited number of qubits, motivating further research into scalable quantum machine learning. The full implementation is made publicly available, and future work will focus on expanding experimental evaluations and publishing additional findings.
Advancements in quantum computing have sparked a systematic evaluation of quantum machine learning (QML) models, focusing on their behavior and performance across diverse learning tasks such as classification, regression, generative modeling, and feature compression. The primary focus is on assessing the performance and behavior of various QML architectures-including Quantum Neural Networks (QNNs), Quantum Support Vector Classifiers (QSVCs), Quantum Generative Adversarial Networks (QGANs), Quantum Convolutional Neural Networks (QCNNs), and Quantum Autoencoders (QAEs)-implemented using Qiskit's machine learning library. This evaluation is conducted using an existing benchmarking framework that ensures consistent experimental conditions and supports flexible evaluation strategies. To provide additional context and insight, these quantum models are compared against classical counterparts using widely adopted implementations. Experiments are carried out on a range of datasets, including structured and low-dimensional examples like Bars and Stripes, as well as more complex datasets like Downscaled MNIST and synthetic Hyperplanes. By providing a standardized evaluation of these models within the Qiskit framework, this study establishes a reproducible performance baseline and highlights key challenges for QML systems in the noisy intermediate-scale quantum (NISQ) era.
Quantum-assisted models can surpass classical counterparts in specific domains by leveraging quantum superposition entanglement for more efficient data representation. This study investigates a Quantum neural network framework executed on various Quantum simulators, including ionq, Cirq, Pennylane and Quantinnum backends. The iris, wine, Indian liver patient and rain datasets are employed as benchmarking datasets to assess the efficacy of quantum-based classification models. The QNN utilises parameterized quantum circuits and quantum feature maps to learn complex patterns in data, optimizing classification decision boundaries through quantum optimization techniques. The study evaluates QNN performance across different quantum hardware platforms, assessing key metrics such as accuracy, recall, F1 score, and precision. Preliminary results indicate that QNN-based classification effectively captures feature dependencies through quantum correlations, offering competitive performance in data classification tasks. These findings contribute to the broader exploration of quantum-enhanced classification and the potential of quantum devices in real-world data analysis.
Quantum machine learning is a new field of study that aims to apply the advantages of quantum computing, such as superposition, which is one of the intriguing properties that could speed up computation, to machine learning. Like classical machine learning models, quantum machine learning models can learn from specific input data and perform specific prediction tasks. However, quantum machine learning models are currently in the form of quantum circuits with parameterizable gates. In this paper, the implementation of a quantum classification model is investigated using the two most well-known quantum machine learning development tools: Cirq with TensorFlow Quantum (TFQ) and Qiskit with PyTorch. The quantum classification circuit is trained to recognize images of handwritten digits using the MNIST dataset. However, because the quantum computer is still in development and is difficult to access and use, quantum simulators are the best alternative for the time being. The performance of the quantum classification model implemented by the two development tools is then compared in this paper experimentally using quantum simulators on two different hardware configurations. In addition, the experimental setups differ in the number of input qubits used to store the image’s classical pixel values. Finally, the wall time and CPU time used to run the training process, as well as classification accuracy, are examined. In terms of accuracy, Qiskit with PyTorch outperforms Cirq with TFQ only in the lower number of qubits but the results are similar in the higher number of qubits. The experiment results in terms of run-time are highly different. From the result, we find that Cirq is noticeably faster than Qiskit in every experiment by at least 90 times.
Quantum neural networks (QNNs) -- parameterised quantum circuits trained to perform pattern recognition and predictive tasks -- represent a prominent but contested direction in quantum machine learning, offering potential parameter efficiency and expressibility benefits over classical neural networks for specific problem structures while suffering from barren plateau trainability limitations at scale. The 2024 issue of this journal established SNAD LSAD-structured ansatz (Silva et al., 2024) as the critical enabler of scalable QNN training, and QHDC (Bianchi et al., 2024) showed that QNN advantage is dataset-dependent. This paper advances the QNN field with the Quantum Neural Network Design and Evaluation (QNNDE) framework, a systematic methodology for QNN architecture design, training optimisation, and performance evaluation across four pattern recognition and prediction task classes: temporal sequence modelling (quantum recurrent circuits), spatial pattern recognition (quantum convolutional circuits), generative modelling (quantum generative adversarial networks), and time-series regression (quantum reservoir computing). QNNDE is evaluated on eight benchmark tasks spanning physics simulation, molecular property prediction, quantum state tomography, financial time-series, medical signal classification, and quantum circuit synthesis on IBM Quantum Eagle (127-qubit) and IonQ Forte (35-qubit). Quantum recurrent circuits achieve 12.4pp accuracy improvement over classical LSTMs for quantum state evolution prediction. Quantum convolutional circuits match classical CNNs on image-derived molecular patterns at 68.4% parameter reduction. Quantum GANs generate molecular conformations with 8.4% lower FID than classical GAN at equal network size. Quantum reservoir computing outperforms classical echo state networks by 18.4% on quantum-process time-series prediction.
Hybrid quantum neural networks (QNNs) pair a classical feature extractor with a variational quantum circuit, and the effective dimension of the circuit's quantum Fisher information matrix (QFIM) is widely used to assess such models. This scalar is often interpreted as an indicator of trainability or accuracy, although these readings are rarely tested under controlled conditions. We therefore investigate what it actually measures. We define the effective dimension as the participation ratio of the QFIM diagonal, which lies between one and the parameter count, and we evaluate it across two image datasets, four frozen backbones including a random-projection control, four circuit depths on eight qubits, and three seeds. During training, the effective dimension remains well below the parameter count, and its correlation with test accuracy ranges from -0.72 to +0.87, changing sign across seeds in half of the configurations. Increasing depth activates a larger share of the circuit directions, from 0.38 to 0.91 of the parameter count, while the initialization gradient variance remains between 1e-5 and 1e-4, so no barren plateau (BP) appears at eight qubits. At a fixed circuit, the four backbones share an initial effective dimension of 6 to 9, even though CIFAR-10 accuracy ranges from 23.14% to 86.44%. These results show that the effective dimension characterizes circuit geometry rather than model quality. We conclude with a practical protocol that assigns each diagnostic to the question it can answer.