QuFeX: Quantum feature extraction module for hybrid quantum-classical deep neural networks
Naman Jain, Amir Kalev · 2025
Read QuFeX: Quantum feature extraction module for hybrid quantum-classical deep neural networks by Naman Jain, Amir Kalev (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-13165.txt.
Introduction
Neural networks that mix quantum and classical parts may sharpen image analysis, but quantum steps are costly to run many times in parallel. Jain and Kalev introduce QuFeX, a quantum module that pulls features out of data in a smaller space. That cuts the number of parallel runs that typical quantum network designs need. The module also slots neatly into ordinary deep networks.
As a test, they place QuFeX at the narrow middle of a U-Net. That design is widely used to outline regions in images, such as organs in medical scans or roads in driving scenes. Their hybrid version, called Qu-Net, outlined regions better than a plain U-Net of the same shape. The authors present QuFeX as a reusable block that could bring quantum help to real tasks that need precise feature extraction.
Catalog record
| CREATOR | Naman Jain, Amir Kalev |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 38 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2501-13165
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
QuFeX: Quantum feature extraction module for hybrid quantum-classical deep neural networks Naman Jain, Amir Kalev Abstract We introduce Quantum Feature Extraction (QuFeX), a novel quantum machine learning module. The proposed module enables feature extraction in a reduced-dimensional space, significantly decreasing the number of parallel evaluations required in typical quantum convolutional neural network architectures. Its design allows seamless integration into deep classical neural networks, making it particularly suitable for hybrid quantum-classical models. As an application of QuFeX, we propose Qu-Net – a hybrid architecture which integrates QuFeX at the bottleneck of a U-Net architecture. The latter is widely used for image segmentation tasks such as medical imaging and autonomous driving. Our numerical analysis indicates that the Qu-Net can achieve superior segmentation performance compared to a U-Net baseline. These results highlight the potential of QuFeX to enhance deep neural networks by leveraging hybrid computational paradigms, providing a path towards a robust framework for real-world applications requiring precise feature extraction. I Introduction Quantum computing has emerged as a computational paradigm with the potential to solve problems intractable for classical computers. The rise of the machine learning (ML) methodology has spurred interest in quantum machine learning (QML) as a potential means to improve classification, regression, and clustering tasks Abohashima et al. 2020 ; Schatzki et al. 2021 . Early on in the development of QML, quantum neural networks (NNs) were proposed Cong et al. 2019 ; Henderson et al. 2019 as an extension of the concept of classical NNs by utilizing quantum circuits as layers of the network. The design of the quantum layers corresponds to different ML architectures, with notable examples including quantum convolutional NN (QCNN) Cong et al. 2019 and Quanvolutional NN (QuanNN) Henderson et al. 2019 . While convolutional NNs (CNNs) – the cornerstone of modern NNs – have achieved significant success across a wide range of ML applications Li et al. 2022a ; Younesi et al. 2024 , their resource requirements, especially for training large models, increase drastically with the size and complexity of data Thompson et al. 2022 ; Hestness et al. 2017 . Quantum NN extensions, such as QCNN and QuanNN, were proposed under the hypothesis that, since quantum computers can process information in fundamentally different ways, quantum NNs may potentially offer greater efficiency compared to their classical counterparts Schuld and Petruccione 2018 ; Dunjko and Briegel 2017 . In parallel to developing quantum NN models, in recent years there has been a growing interest in creating a hybrid quantum-classical model where quantum and classical NNs work in tandem Zaman et al. 2024 ; Hafeez et al. 2024 ; Zhang et al. 2024 ; Arthur and Date 2022 . This division of computing resources allows leveraging the strengths of both classical and quantum computing while managing the limitations of near-term quantum devices. Hybrid models can be particularly promising in domains such as image processing and generative modeling, where quantum circuits may offer advantages in high-dimensional feature spaces, while classical systems handle more routine tasks. Thus, the hybrid quantum-classical ML approach provides a feasible pathway toward enhancing ML tasks under the current quantum technological constraints. A typical hybrid model includes classical layers for pre-processing data, followed by quantum processing layers (circuits), and concludes with classical post-processing steps. Nevertheless, various strategies have been proposed for integrating quantum layers into classical deep NN architectures. For example, in image classification tasks, quantum layers are often positioned near the end of the hybrid network to generate the final output label Senokosov et al. 2024 , while some approaches involve encoding data through quantum layers at the beginning of the network Li et al. 2022b . The design of the quantum layer itself varies widely as well, with notable examples including QCNN Cong et al. 2019 and QuanNN Henderson et al. 2019 . These variations underscore the need for a comprehensive study of optimal quantum layer placement in hybrid networks, as well as a systematic approach to constructing effective quantum layers. In this work, we take a step forward in these two directions. We introduce a novel quantum learning architecture, which we term quantum feature extraction, or QuFeX for short, which combines and leverages techniques from QCNN Cong et al. 2019 and QuanNN Henderson et al. 2019 . At a high-level it integrates the data analysis structure of QCNN, which was shown to excel in classification tasks Elhag et al. 2024 ; Farhi and Neven 2018 ; Hur et al. 2021 , with the data pipeline of QuanNN which is well suited for local transformation of the input data Henderson et al. 2019 . This combination allows us to design a quantum layer that can extract features from the data in a reduced dimensional representation of it. This in turn results in a reduced number of parallel quantum circuit evaluations compared to QuanNN. Thus, one of the main advantages of the QuFeX design is in its ability to be integrated in a hybrid fashion with classical layers, which makes it an excellent candidate for integration in deep classical NNs. As a proof of concept, we propose a novel hybrid quantum-classical model, a Qu-Net, were the QuFeX is integrated at the bottleneck of a classical U-Net model. U-Nets are a type of convolutional neural network architecture primarily designed for image segmentation tasks Ronneberger et al. 2015 . They are widely used in applications such as medical imaging Ronneberger et al. 2015 ; Yuan and Cheng 2024 , autonomous driving Iglovikov and Shvets 2018 , and satellite image analysis Bastani et al. 2018 , where precise pixel-level predictions are essential. Our findings thus provide numerical evidence that with a careful design of quantum learning circuits, and with a strategic placement of those within deep NNs we can achieve quantum-classical hybrid learning models that leverage the strengths of both computing paradigms. Our demonstrations represent one of the first applications of quantum-enhanced deep NNs to real-world image segmentation tasks. The remainder of the paper is organized as follows. In Section II , we first present a brief overview of CNNs and their quantum counterpart architectures, setting the stage for the introduction of the QuFeX module. In Section III , we introduce the Qu-Net architecture utilizing the QuFeX at the bottleneck of the network. We numerically test and analyze the performance of the Qu-Net , for various model sizes, compared to an all-classical U-Net baseline. Finally, we offer conclusions and outlook in Section IV . II The QuFeX architecture Figure 1: Typical CNN architecture . The network consists of alternating convolutional and pooling layers for feature extraction, followed by fully connected layers usually for classification. The figure also illustrates the convolution operation, showing how a filter interacts with the input data to produce feature maps. Setting the stage for the proposed quantum architecture, we first highlight a few key components of classical CNNs, QCNNs, and QuanNNs that are relevant to our work. II.1 Convolutional neural networks and their quantum counterparts At its core, a CNN is composed of a hierarchal series of convolutional layers, which are the fundamental building blocks of modern classical NNs. As illustrated in Fig. 1 each convolutional layer computes new feature values from a weighted linear combination of nearby ones in the preceding layer. The