Insights on Building Quantum Convolutional Neural Networks

In this paper, we propose a multiscale entanglement renormalization ansatz (MERA) feature extraction method based on a novel quantum convolutional neural network (QCNN) for binary scanning tunneling microscopy (STM) image classification. We design QCNN quantum circuits for state preparation, quantum convolution, and quantum pooling in the TensorFlow quantum framework and compare the performance of QCNN classifier and two hybrid quantum-classical QCNN models. Adversarial attacks are considered as a type of interpretable method to evaluate the robustness of QCNN models. The similarity between the pixels of image bitplane slicing and Ising phase transition opens up new ways for exploring classification performance enhancement by QCNN classifiers. Classification performance of different bitplanes of QCNN also shows that they can robustly resist adversarial attacks such as FGSM, CW, JSMA, and DEEPFOOL.

The paper analyses and investigates the usage of quantum convolutional neural networks in technical, natural, and socio-economic systems. Quantum convolutional neural networks are a novel approach to information processing that is based on the principles of quantum mechanics and artificial intelligence. In technical systems, the potential of using quantum convolutional neural networks for solving complex tasks such as image processing, machine learning, and prediction has been explored. The results have shown that quantum convolutional neural networks can provide more accurate and faster computations compared to classical neural networks. In natural systems, research has been conducted on the use of quantum convolutional neural networks for modeling and predicting complex natural processes. Their effectiveness in understanding genetic data, studying complex molecular structures, and analyzing ecological systems has been investigated. It has been found that quantum convolutional neural networks can deliver more precise and rapid results compared to conventional data processing methods. In socio-economic systems, the possibilities of employing quantum convolutional neural networks for social network analysis, financial market forecasting, and resource management have been studied. The application of quantum convolutional neural networks has the potential to enhance prediction accuracy and facilitate more effective decision-making in socio-economic systems. The research findings confirm that quantum convolutional neural networks have the potential to be utilized in various domains, including technical, natural, and socio-economic systems. They can achieve higher accuracy, processing speed, and predictive capabilities compared to traditional methods.

Neural network-based algorithms have garnered considerable attention in condensed matter physics for their ability to learn complex patterns from very high dimensional data sets towards classifying complex long-range patterns of entanglement and correlations in many-body quantum systems. Small-scale quantum computers are already showing potential gains in learning tasks on large quantum and very large classical data sets. A particularly interesting class of algorithms, the quantum convolutional neural networks (QCNN) could learn features of a quantum data set by performing a binary classification task on a nontrivial phase of quantum matter. Inspired by this promise, we present a generalization of QCNN, the branching quantum convolutional neural network, or bQCNN, with substantially higher expressibility. A key feature of bQCNN is that it leverages mid-circuit (intermediate) measurement results, realizable on current trapped-ion systems, obtained in pooling layers to determine which sets of parameters will be used in the subsequent convolutional layers of the circuit. This results in a branching structure, which allows for a greater number of trainable variational parameters in a given circuit depth. This is of particular use on current-day NISQ devices, where circuit depth is limited by gate noise. We present an overview of the ansatz structure and scaling, and provide evidence of its enhanced expressibility compared to QCNN. Using artificially-constructed large data sets of training states as a proof-of-concept we demonstrate the existence of training tasks in which bQCNN far outperforms an ordinary QCNN. Finally, we present future directions where the classical branching structure and increased density of trainable parameters in bQCNN would be particularly valuable.

This paper presents a quantum convolutional neural network (QCNN) for the classification of high energy physics events. The proposed model is tested using a simulated dataset from the Deep Underground Neutrino Experiment. The proposed quantum architecture demonstrates an advantage of learning faster than the classical convolutional neural networks (CNNs) under a similar number of parameters. In addition to the faster convergence, the QCNN achieves a greater test accuracy compared to CNNs. Based on our results from numerical simulations, it is a promising direction to apply QCNN and other quantum machine learning models to high energy physics and other scientific fields.

- We present a comprehensive study exploring quantum machine learning (QML) approaches for plant disease detection from leaf images and compare them against well-established classical counterparts. Specifically, we implement and analyze Quantum Support Vector Machines (QSVMs) vs classical SVMs, and Quantum Convolutional Neural Networks (QCNNs) vs classical CNNs. Using the widely used PlantVillage and complementary field datasets, we describe image preprocessing, classical baseline architectures, quantum data-encoding strategies, circuit-level QSVM and QCNN designs for near-term quantum devices, and hybrid training procedures. Where possible, we review literature-reported performance and propose a reproducible experimental pipeline for empirical evaluation on simulated/noisy quantum backends. We discuss expected strengths and limitations of quantum approaches (expressivity, kernel advantages, resource constraints), provide detailed evaluation metrics and ablations, and propose directions for real-device experiments and field deployment. Key takeaways: QSVM/quantum-kernel methods can provide superior separability on certain feature maps and small-to-medium-sized datasets, while QCNNs show promise as compact feature extractors for hybrid pipelines — but both approaches currently require careful circuit design and error-mitigation to outperform well-tuned classical models in realistic field settings. Key Words: QCNN, Plant Disease, SVM, CNN, QSVM

In recent years, Variational Quantum Circuit (VQC)s have been widely explored to advance quantum circuits against classic models on various domains, such as quantum chemistry and quantum machine learning. Similar to classic machine-learning models, VQCs can be optimized through gradient-based approaches. However, the gradient of VQCs may dramatically vanish during the optimization process as the number of qubits or layers increases. This issue, known as Barren Plateau (BP)s, seriously hinders the scaling of VQCs on large datasets. To mitigate the exponential gradient vanishing, extensive efforts have been devoted to tackling this issue through different methods. Some BP mitigation strategies include VQC parameter initialization and more complex VQC architectures such as Quantum Neural Network (QNN)s and Quantum Convolutional Neural Network (QCNN). In this thesis, we aim to investigate several circuit components through the lens of parameter initialization techniques. We study the performance and gradient variance of the different quantum circuit structures in combination with various initialization methods. In most existing studies, researchers typically focus on improving or understanding one aspect of the VQC approach. Instead, we plan to investigate how the two components of the VQC behave as we interchange the different variables of the circuit. This investigation will provide a better understanding of how the VQC structures behave under different initialization strategies for binary classification problems.

Software vulnerabilities pose a significant threat to system security, necessitating effective automatic detection methods. Current techniques face challenges such as dependency issues, language bias, and coarse detection granularity. This study presents a novel deep learning-based vulnerability detection system for Java code. Leveraging hybrid feature extraction through graph and sequence-based techniques enhances semantic and syntactic understanding. The system utilizes control flow graphs (CFG), abstract syntax trees (AST), program dependencies (PD), and greedy longest-match first vectorization for graph representation. A hybrid neural network (GCN-RFEMLP) and the pre-trained CodeBERT model extract features, feeding them into a quantum convolutional neural network with self-attentive pooling. The system addresses issues like long-term information dependency and coarse detection granularity, employing intermediate code representation and inter-procedural slice code. To mitigate language bias, a benchmark software assurance reference dataset is employed. Evaluations demonstrate the system's superiority, achieving 99.2% accuracy in detecting vulnerabilities, outperforming benchmark methods. The proposed approach comprehensively addresses vulnerabilities, including improper input validation, missing authorizations, buffer overflow, cross-site scripting, and SQL injection attacks listed by common weakness enumeration (CWE).

The rapid evolution of Artificial Intelligence has enabled significant progress in image classification, with emerging approaches extending traditional deep learning paradigms. This article presents an extended version of a paper originally introduced at ICCSA 2025, providing a broader comparative analysis of classical, spline-based, and quantum machine learning architectures. The study evaluates Convolutional Neural Networks (CNNs), Kolmogorov–Arnold Networks (KANs), Convolutional KANs (CKANs), and Quantum Convolutional Neural Networks (QCNNs) on the Labeled Faces in the Wild dataset. In addition to these baselines, two novel architectures are introduced: a fully quantum Kolmogorov–Arnold model (F-QKAN) and a hybrid KAN–Quantum network (H-QKAN) that combines spline-based feature extraction with variational quantum classification. Rather than targeting state-of-the-art performance, the evaluation focuses on analyzing the behaviour of these architectures in terms of accuracy, computational efficiency, and interpretability under a unified experimental protocol. Results show that the fully quantum F-QKAN achieves a test accuracy above 80%. The hybrid H-QKAN obtains the best overall performance, exceeding 92% accuracy with rapid convergence and stable training dynamics. Classical CNNs models remain state-of-the-art in terms of predictive performance, whereas CKANs offer a favorable balance between accuracy and efficiency. QCNNs show potential in ideal noise-free settings but are significantly affected by realistic noise conditions, motivating further investigation into hybrid quantum–classical designs.

Skin cancer, in particular melanoma, is still one of the most significant contributing factors to cancer-related deaths globally, indicating the urgency of reliable diagnostic tools. Early discovery of skin lesions can radically increase patient survival rates, but manual diagnosis is not an easy task, and it is time-consuming as well. This paper uses Quantum Convolutional Neural Networks (QCNN), a hybrid quantum-classical model for the skin lesion classification task as being benign or malignant. The model intertwines quantum feature extraction using an amplitude encoding and parameterized quantum circuit with classical convolutional neural network (CNN) layers analyzing spatial features. Advanced pre-processing techniques and Synthetic Minority Over-sampling Techniques (SMOTE) address the class imbalance, as federated learning ensures decentralized nodes’ privacy-preserving training. The explainability of the model predictions is heightened through Quantum SHAP (Shapley Additive Explanations) and Gradient-Weighted Class Activation Mapping (Grad-CAM). This work evaluates QCNN on the Melanoma Skin Cancer Dataset of 10,600 images, achieving 96.87% training accuracy and a validation accuracy of 85% while surpassing classical benchmarks in accuracy, efficiency, and robustness. Therefore, this quantum-classical model has highlighted potential applications of medical image diagnostics through quantum-classical models

This research work on the well-known WM811K wafer map dataset intends to integrate the insights provided by wafer manufacturing techniques and the computing power of quantum convolutional neural networks (QCNN) approach. The work begins with demonstrating high accuracy (𝑹 𝟐 =99.32%) of the classical convolutional neural network (CCNN) consisting of three consecutive convolution layers on the WM-811K dataset. In later part of the paper, optimized for multi-class classification, the applied QCNN model consisting of four qubits used for convolutional and pooling layers exhibit 99.38% accuracy. The QCNN model establishes supremacy over CCNN model by reducing time requirement to train the model for classifying defects on the wafer images.

Plants are susceptible to numerous illnesses during their growth stages, rendering early diagnosis a significant challenge in contemporary agriculture. Not being able to find infections early on can greatly lower agricultural yields and hurt farmers' incomes. There are several new methods that use Machine Learning and Deep Learning to solve this problem; however they generally have low classification accuracy or need millions of trainable parameters, which makes them less efficient. To overcome these limitations, this study proposes a Quantum Convolutional Neural Network (QCNN) model for the automated detection of plant diseases. To the best of our knowledge, there is currently no literature that has investigated QCNNs for this purpose. This work employs the proposed approach to detect Bacterial Spot disease in peach plants utilizing leaf photos from the publicly accessible Plant Village dataset. The system only needs 9,914 trainable parameters to work, and it gets 94.96 % accuracy. These results show that QCNNs are good at predicting plant diseases and could be useful in other areas of agriculture as well.

In this study, we extract agricultural (agri-) workers’ physical acceleration timeline data and execute quantum neural network-based deep learning for data classification. Although various approaches have been implemented globally for indoor and outdoor agri-working sites, there is still scope for improvement. Therefore, in this study, we adapt these approaches mainly for automated, high-tech agri-sites and apply the quantum convolutional neural network (QCNN) deep learning-based method to qualitatively demonstrate the classification of physical workers’ timeline datasets. For our dataset, our subjects were six experienced and six completely inexperienced male agri-manual workers. The target task was to cultivate a farmland using a simple traditional Japanese hoe in a semi-crouching position. We captured the subjects’ acceleration data using an integrated multi-sensor module mounted on a wooden lilt 150 mm from the gripping position of the dominant hand. We developed and used Python (ver. 3.9) and recent distributed libraries for quantum computation and distributed concerned libraries. For data classification, we successively executed three patterns using QCNN, a category of quantum deep learning (QDL)-based deep learning, and evaluated the systems by obtaining loss, validation accuracy, and final epochs’ elapsed time data. These methods of analyzing digital data are expected to find practical applications and provide key suggestions for improving daily task efficiency.

While Quantum Convolutional Neural Networks (QCNNs) show great potential for image classification tasks, their application is hampered by the lack of systematic guidance for hyperparameter configuration. The selection process, which is often a matter of trial-and-error, leads to suboptimal and inefficient model performance. This research addresses this issue by constructing an experimental framework to perform a comprehensive sensitivity analysis of QCNN hyperparameters. Using a constrained grid search method, thousands of model configurations were evaluated on the MNIST, Fashion-MNIST, and CIFAR-10 datasets under two scenarios: an ideal environment and an environment with a realistic quantum noise model. Analysis using non-parametric statistical tests reveals that nearly all tested hyperparameter categories, including the optimizer, embedding method, ansatz architecture, reduction method, and cost function, have a statistically significant impact on performance. The choice of optimizer and data embedding method were found to be the two most crucial factors. A key finding of this study is that the choice of hyperparameters actively and significantly determines the model's robustness to noise, revealing a clear trade-off between optimal performance in ideal conditions and resilience in noisy ones. This research provides a practical, evidence-based guide that can help researchers prioritize tuning to achieve more effective and efficient QCNN performance.

Animal welfare within livestock production is growing in importance to consumers, challenging livestock producers to employ practices that promote animal welfare and increase transparency while meeting output goals. Therefore, livestock production systems may benefit from using digital technologies to increase production efficiency and minimize physical handling of animals, which may enhance consumer perceptions of welfare. Artificial intelligence (AI) technologies that leverage computer vision to monitor cattle are emerging, but research is limited. Accordingly, the objective of this research was to develop three AI models using practical technologies with Deep Convolutional Neural Networks (DCNN) to predict cattle behavior, movement, and emotion. Procedures involving animals were approved by Texas State University Institutional Animal Care Use Committee (#9298). Groups (n~2-7) of cattle were housed on pastured paddocks and continuously recorded with nine stationary cameras. Cattle were phenotypically and physiologically heterogeneous and rotated to provide robust footage for model development via inter-animal variation. Captured raw footage (~132 h) was curated, edited into 5-30-second clips using VideoLAN, and labeled with vague descriptions of the cattle’s activity. Labeled clips were then annotated using Computer Vision Annotation Tool (CVAT) software by identifying cattle with bounding boxes and labeling them as “bovine.” Specific attributes of “behavior” (e.g., grazing, playing, dominance assertion), “movement” (e.g., walking, standing, lying), and “emotion” (e.g., calm, excitement, curiosity) were added to the bovine label. The computer vision model You Only Look Once (YOLO) was used to identify cattle and classify their behavior, movement, and emotion. Footage was split into segments: images used for training (14,340 frames; 89%) and validation (1,730 frames; 11%). Results indicate that the cattle behavior model had optimal levels of precision and recall for “grazing, “hay feeding,” and “normal” behaviors, collectively known as “all classes” (78%), with a mean average precision (mAP) of 0.5, and 78.0% confidence predicting “all classes.” The cattle movement model had optimal levels of precision and recall for “walking,” “standing up,” “standing,” and “lying” movements, collectively known as “all classes” (59.7%) with a mAP of 0.5, and 58.0% confidence predicting “all classes.” The cattle emotion model had optimal levels of precision and recall occurring for the “calm” emotion, collectively known as “all classes” (96.1%) with a mAP of 0.5, and 90.0% confidence predicting “all classes.” Our data indicate that the three developed models can accurately detect cattle and predict behaviors, movements, and emotions using practical technological resources. The models will be trained on a larger dataset before testing. While establishing an AI-enabled system for livestock monitoring can enhance production and animal welfare, prior to implementing such technologies, future research should develop and test the ability of the system to detect specific signs of stress, animal health, and desired behaviors.

Computer vision systems are increasingly expanding their application in visual data analysis. Model training methods are undergoing the greatest development and improvement as the results of this stage significantly affect the final classification of objects and the interpretation of input information. Typically, computer vision systems use convolutional neural networks for training (Convolution Neural Network, CNN). The disadvantages of such systems are significant limitations in cross-modal learning, multimodality implementation, labeling of large amounts of data, etc. One of the ways to overcome these problems is to use Vision Transformers (ViT), which, compared to classical CNNs, have higher performance due to reduced inductive biases and high parallel computing efficiency. Introducing Self-Supervised Learning (SSL) technologies can significantly reduce the dependence on manually labeled data, contributing to the formation of generalized representations of images. Cross-Modal Learning (CML) expands the possibilities of processing them by combining data of different types. The development of the new approach, combined with the capabilities of cross-modal learning and self-learning in ViT in a single architecture, will ensure adaptability, efficiency, and system scalability in various applications. The research aims to provide a detailed overview of ViTs, approaches to their architecture, and methods for improving their efficiency. The mathematical foundations of the key concepts of ViT, cross-modal learning and self-learning, the main modifications of ViT, and their integration with SSL and CML technologies are considered. A comparison of methods using characteristics, performance, and efficiency is provided. The key challenges and prospects facing researchers and developers while creating universal models in computer vision are outlined. ViTs change computer vision by capturing global dependencies on images. Despite some challenges, ViTs provide excellent scalability and performance for large datasets. The active search for methods to overcome their limitations makes ViTs a key tool for improving image classification, object detection, and other computer vision tasks.