Hybrid Quantum-Classical Learning for Multiclass Image Classification

Shuchismita Anwar, Sowmitra Das, Muhammad Iqbal Hossain, Jishnu Mahmud · 2025

Read Hybrid Quantum-Classical Learning for Multiclass Image Classification by Shuchismita Anwar, Sowmitra Das, Muhammad Iqbal Hossain, Jishnu Mahmud (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-2508-18161.txt.

Introduction

Shuchismita Anwar and co-authors rescue thrown-away states in quantum image sorters in a 2025 paper. Today's quantum nets shrink images by pooling, and the dropped units still hold links to the kept ones, so real data is lost. The team catches the dropped states and feeds them to a small classic net of their own.

The kept states go to a twin classic net, and the two votes merge before the final call. Training tunes both sides at once with one loss. On digit, fashion, and organ scan sets, the blend beat lean rivals of the same size. Saving dropped states looks like a rich path for mixed quantum tools.

Catalog record

CREATORShuchismita Anwar, Sowmitra Das, Muhammad Iqbal Hossain, Jishnu Mahmud
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE48 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-18161

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Hybrid Quantum-Classical Learning for Multiclass Image Classification Shuchismita Anwar, Sowmitra Das, Muhammad Iqbal Hossain, Jishnu Mahmud Abstract This study explores the challenge of improving multiclass image classification through quantum machine‑learning techniques. It explores how the discarded qubit states of Noisy Intermediate‑Scale Quantum (NISQ) quantum convolutional neural networks (QCNNs) can be leveraged alongside a classical classifier to improve classification performance. Current QCNNs discard qubit states after pooling; yet, unlike classical pooling, these qubits often remain entangled with the retained ones, meaning valuable correlated information is lost. We experiment with recycling this information and combining it with the conventional measurements from the retained qubits. Accordingly, we propose a hybrid quantum–classical architecture that couples a modified QCNN with fully connected classical layers. Two shallow fully connected (FC) heads separately process measurements from retained and discarded qubits, whose outputs are ensembled before a final classification layer. Joint optimisation with a classical cross‑entropy loss allows both quantum and classical parameters to adapt coherently. The method outperforms comparable lightweight models on MNIST, Fashion‑MNIST and OrganAMNIST. These results indicate that reusing discarded qubit information is a promising approach for future hybrid quantum–classical models and may extend to tasks beyond image classification. 0 0 footnotetext: This is a preprint. I Introduction Quantum machine learning combines the representational power of artificial intelligence with the parallelism and entanglement of quantum computing, offering a practical route to tackle computationally intensive tasks such as multiclass image recognition Henderson et al. 2019 ; Riaz et al. 2023 ; Ciliberto et al. 2018 . While classical convolutional neural networks (CNNs) have long dominated visual recognition, their accuracy gains come at the cost of ever-growing depth, parameter counts, and energy consumption Kim et al. 2023 ; Li et al. 2020 ; Vatan and Williams 2004a ; Chalumuri et al. 2021a . Quantum Convolutional Neural Networks (QCNNs) mitigate these overheads by executing convolution and pooling through unitary operations, achieving comparable or occasionally superior performance with far fewer parameters and shallower circuits, an advantage that is especially relevant for today’s noisy intermediate-scale quantum (NISQ) hardware Henderson et al. 2019 ; Kim et al. 2023 ; Cong et al. 2019 ; Gianani et al. ; Liang et al. 2021 ; Preskill 2018 ; Bharti et al. 2022 ; Easom-McCaldin et al. 2024 ; Hur et al. 2022 ; Chandra et al. 2022 ; Wei et al. 2022 ; Abbas et al. 2021 . Empirical studies already report strong QCNN accuracies on benchmarks such as MNIST, Fashion-MNIST, and various medical-imaging datasets Henderson et al. 2019 ; Kim et al. 2023 ; Easom-McCaldin et al. 2024 ; Hur et al. 2022 ; Chandra et al. 2022 . For example, Henderson et al. surpassed 88% on a brain-disease diagnostic task despite using only tens of quantum parameters Bokhan et al. 2022 . Hybrid quantum–classical models (HQCNNs) extend this promise, delivering competitive results on tabular problems like Iris and on Wi-Fi-based indoor localisation, where the quantum front-end extracts low-dimensional yet highly separable features Hur et al. 2022 ; Chalumuri et al. 2021b . Recent work has moved beyond raw circuit design toward specialised quantum embeddings, lightweight unitary ansätze, and problem-specific classifiers, often outperforming classical networks with an order of magnitude fewer parameters Riaz et al. 2023 ; Li et al. 2020 ; Chalumuri et al. 2021a ; Gianani et al. ; Chandra et al. 2022 ; Chalumuri et al. 2021b ; Lloyd et al. 2020 ; Jain et al. 2020 ; Wang et al. 2021 ; Hur et al. 2024 . These gains are frequently attributed to embeddings that maximise class separation in Hilbert space, thereby reducing the burden on the subsequent classifier Riaz et al. 2023 ; Gianani et al. ; Chandra et al. 2022 ; Wei et al. 2022 ; Chalumuri et al. 2021b ; Hur et al. 2024 . Parallel advances in Quantum Support Vector Machines highlight the same trend: carefully engineered quantum feature maps can deliver superior decision boundaries with logarithmic complexity in the data dimension Rebentrost et al. 2014 ; Lloyd et al. 2014 . Despite promising results; 98.97% on MNIST and 91.40% on GTSRB for Quantum Deep CNNs Li et al. 2020 , for example - scale and noise resilience remain open challenges as datasets grow in size and complexity Ciliberto et al. 2018 ; Easom-McCaldin et al. 2024 ; Di et al. 2023 ; Liu et al. 2021 ; LaRose and Coyle 2020 ; Zhao et al. 2019 ; Farhi and Neven 2018 ; Schuld et al. 2020 ; Haah 2019 ; Kerenidis and Prakash 2016 ; Kerenidis and Luongo 2020 ; Dervovic et al. 2018 . Novel variants such as the Amplitude-Transformed QCNN (ATQCNN) improve data-efficiency for domain-specific inputs, yet still struggle on heterogeneous datasets when the parameter budget is tightly capped Di et al. 2023 . Tensor-network methods achieve above-98% accuracy in binary settings but often falter on multiclass tasks due to limited expressive power once the bond dimension is fixed Wang et al. 2021 . Other hybrid schemes have demonstrated utility in medical imaging; for example, Liang et al. for Alzheimer’s detection and Lusnig et al. for federated liver analysis yet they too inherit the same bottleneck as they scale beyond binary or small-class problems Liang et al. 2021 ; Lusnig et al. 2023 ; Akpinar 2023 . Traditional and hybrid QCNNs achieve dimensionality reduction through pooling layers by omitting qubits, which causes quantum information loss and compromised performance, especially in NISQ devices with scarce qubits Wei et al. 2022 ; Easom-McCaldin et al. 2024 ; Lusnig et al. 2023 ; Hur et al. 2024 ; García-Ripoll 2021 . However, previous research has not sufficiently addressed the limitations arising from discarded quantum information during pooling, nor has it explored fully exploiting quantum-classical interactions to enhance classification performance. This gap significantly limits QCNN potential for complex real-world data applications. To address these challenges, in this work, we suggest a novel hybrid quantum-classical architecture with improved classification performance, conserving quantum information, and increased resource efficiency. The following work proposes a novel Hybrid Quantum-Classical Architecture that addresses a number of challenges with several important contributions: • Proposed Scaling Method : The architecture proposes to extend the range of quantum measurement values which optimally maps quantum data onto the classical values. The scaling of the measurements ensures more precise and efficient quantum operations which raises the accuracy of the hybrid model during the classification process. • Utilization of Discarded Qubits from the Pooling Layer : Instead of discarding the qubit states completely from the pooling layer like traditional QCNN’s, the proposed architecture reuses such discarded qubits. Further processing extracts useful quantum information from them; it also adds it to the final computations and thereby improving the performance in classification with limited computational overhead. The method avoids loss of information as well as strengthens the capability of the model to handle complicated data. • Classical Layers for Joint Optimization : The scaled values of discarded qubits and the retained qubits undergo classical layers separately. After that, their outputs get dot-multiplied. This captures more of the quantum features that will contribute in the overall joint optimization of the model. These classical layers back-propagate to dynamically adjust parameters on the combined quantum-classical data. The classical layers further optimize decision

Read the complete text