Abstract We present QNN-3L, a quantum neural network architecture based on three-level systems (qutrits) with longitudinal Ising-type couplings on a one-dimensional lattice. Starting from an explicit phenomenological multilevel Hamiltonian that includes local anharmonicity, diagonal couplings and resonant control pulses, we derive native selective rotations on computational subspaces and a nearest-neighbour controlled-\((Z)\) primitive. We prove that these native operations generate a subgroup dense in \((\mathrm{SU}(3^N))\), providing approximate universality. To introduce effective nonlinearity while preserving coherent unitary propagation, we propose a hybrid architecture that uses a multi-subspace feature map: classical inputs are encoded via sequential rotations on two independent qutrit subspaces, whose non-commutativity enriches the implicit quantum kernel. We perform ablation studies and kernel diagnostics showing that the multi-subspace encoding yields improved kernel separation and more stable training than single-subspace alternatives. Numerical benchmarks for \((N=3)\)–6 qutrits (dimension \((\dim\mathcal{H}\leq729)\)) demonstrate comparable expressibility at depths \((L\ge3)\) and enhanced shallow-circuit entangling power relative to equivalent qubit circuits, and validate proof-of-concept supervised learning on a canonical nonlinear task. MSC Classification: 03.67.Ac , 03.67.Lx , 07.05.Mh
Abstract We present QNN-3L, a quantum neural network architecture based on three-level systems (qutrits) with longitudinal Ising-type couplings on a one-dimensional lattice. Starting from an explicit phenomenological multilevel Hamiltonian that includes local anharmonicity, diagonal couplings and resonant control pulses, we derive native selective rotations on computational subspaces and a nearest-neighbour controlled-((Z)) primitive. We prove that these native operations generate a subgroup dense in ((\mathrm{SU}(3^N))), providing approximate universality. To introduce effective nonlinearity while preserving coherent unitary propagation, we propose a hybrid architecture that uses a multi-subspace feature map: classical inputs are encoded via sequential rotations on two independent qutrit subspaces, whose non-commutativity enriches the implicit quantum kernel. We perform ablation studies and kernel diagnostics showing that the multi-subspace encoding yields improved kernel separation and more stable training than single-subspace alternatives. Numerical benchmarks for ((N=3))–6 qutrits (dimension ((\dim\mathcal{H}\leq729))) demonstrate comparable expressibility at depths ((L\ge3)) and enhanced shallow-circuit entangling power relative to equivalent qubit circuits, and validate proof-of-concept supervised learning on a canonical nonlinear task. MSC Classification: 03.67.Ac , 03.67.Lx , 07.05.Mh
Neural quantum kernels (NQKs) construct quantum kernels by pretraining a quantum neural network (QNN) and subsequently reusing the trained circuit as a task-adapted embedding. Extending this framework to qudits, with local unitaries in SU(d), provides a natural route to richer data embeddings through the increased local degrees of freedom and a direct interface for multiclass classification via intrinsically multi-level quantum systems. In this work, focusing on qutrits (d=3), we extend NQKs to the qudit setting and perform a systematic study of key design choices, including the number of encoded features, the number of qutrits, the kernel construction (1-to-n and n-to-n), and the parameterization of SU(3) unitaries. Across binary and three-class tasks on four benchmark datasets, qutrit NQKs improve over the corresponding QNN baselines in nearly all settings considered and can benefit from scaling both the feature budget and the system size, although the magnitude of these gains may saturate, is dataset-dependent, and depends on the chosen parameterization. In particular, an ablation over SU(3) parameterizations shows that the unitary representation can substantially impact both optimization behaviour and classifier performance. These findings highlight the potential of qudit-based quantum models not only as a straightforward generalization of qubit-based architectures, but also as a promising means to better exploit complex data structures in quantum machine learning.
In this work, we present a training method for a Fully Quantum Neural Network (FQNN) based entirely on quantum circuits. The model processes data exclusively through quantum operations, without incorporating classical neural network layers. In the proposed architecture, the roles of classical neurons and weights are assumed, respectively, by qubits and parameterized quantum gates: input features are encoded into quantum states of qubits, while the network weights correspond to the rotation angles of quantum gates that govern the system’s state evolution. The optimization of gate parameters is performed using directional gradient estimation, where gradients are numerically approximated via finite differences, eliminating the need for analytic derivation. The training objective is defined as the quantum-state fidelity, which measures the similarity between the network’s output state and a reference state representing the correct class. Experiments were conducted using the Qiskit AerSimulator, which allows for the accurate simulation of quantum circuits on a classical computer. The proposed approach was applied to the classification of the Iris dataset. The experimental results demonstrate that the FQNN is capable of effectively learning to distinguish between classes based on input features, achieving stable test accuracy across runs. These findings confirm the feasibility of constructing fully quantum classifiers without relying on hybrid quantum—classical architectures. The FQNN architecture consists of multiple quantum layers, each incorporating parameterized rotation operations and entanglement between qubits. The number of layers is determined by the ratio of quantum parameters (weights) to the number of input features. Each layer functions analogously to a hidden layer in a classical neural network, transforming the quantum-state space into a richer feature representation through controlled quantum operations. As a result, the network is capable of dynamically modeling dependencies among input features without the use of classical activation functions.
Quantum neural networks (QNNs) have recently attracted remarkable attention as a potential alternative to classical and deep neural network (NN) models particularly under constraints imposed by noisy intermediate-scale quantum (NISQ) devices, which support only shallow circuits with few quantum bits (qubits) and limited variational depth. However, applying QNNs to natural language processing (NLP) remains challenging because text representations are typically high-dimensional, while NISQ classifiers can only encode a small number of features at a time, requiring classical compression to enable stable training and flexible circuit configuration. In addition, many existing studies integrate QNNs into rich-text representation pipelines, where the quantum circuit is often implemented as a variational quantum circuit (VQC). Here, the VQC acts as an additional component within a larger model, and it remains unclear how the quantum component improves the classification task. Therefore, in this work, we propose a fully reproducible and compact QNN pipeline that provides a controlled setting for studying the contribution of the VQC to text classification. The model is trained end-to-end with a lightweight classical compression layer, and the overall study is supported by systematic ablation analyses of the compression layer, the classifier heads under a shared frozen low-dimensional representation, and quantum circuit structure. Our model uses the term frequency-inverse document frequency (TF-IDF) features with linear compression followed by a shallow four-qubit, two-layer VQC, implemented in PyTorch and PennyLane for end-to-end training. It is evaluated on five benchmark text classification datasets against classical baselines, including logistic regression and multilayer perceptron (MLP) models. Experimental results show that the proposed QNN achieves a competitive performance across all evaluated datasets, yielding the best results on three of the five benchmark datasets.
Quantum neural networks (QNNs) integrate the advantages of quantum information with the formidable capabilities of classical neural networks, exhibiting notable potential in image classification tasks. However, the performance of current QNNs in processing multi-class image classification tasks still leaves room for improvement. In this paper, we propose three QNN architectures to enhance the accuracy of multi-class image classification by utilizing techniques of overparameterization, dense pooling, and local feature extraction. These networks are capable of mapping conventional data to the amplitudes of quantum states via amplitude encoding, eschewing the need for traditional dimensionality reduction techniques. Following processing with parameterized quantum circuits (PQCs), we conduct quantum measurements to extract feature information. We validate the proposed architectures using the MNIST dataset for eight-class image classification problems. The experimental results unequivocally demonstrate that the three techniques we employed significantly enhance the accuracy of QNNs in multi-class image classification tasks.
Abstract Quantum Neural Networks (QNNs) present a promising research direction in image recognition, demonstrating significant potential for various applications. However, QNNs with a fixed circuit topology structure are prone to encountering the barren plateaus during training when processing exponential number of qubits, which limit their performance and scalability. In this paper, a novel Hybrid Quantum-Classical Inception-inspired Convolutional Neural Networks (HQCNN) is proposed. The HQCNN enhances feature learning capability and mitigates the barren plateau issue by integrating quantum convolutional kernels with diverse circuit topologies across multiple channels. The architecture employs an amplitude encoding layer to map multi-channel classical data into quantum states, followed by diverse quantum convolutional kernels that extract feature representations. Furthermore, Hadamard gates are used to intergrate multi-channel features and incorporate quantum pooling to reduce feature dimensionality, with a fully connected layer mapping the quantum-enhancing features into class labels for classification tasks. Finally, by analyzing the influence of the circuit structures and quantum gate combinations on the models' performance, the optimal HQCNN architecture is constructed. Experimental results demonstrate that HQCNN achieves 98.5% accuracy on the MNIST-4 classification task and 92.5% accuracy on the MNIST-10 classification task, which indicates that the proposed model outperforms state-of-the-art HQCNN models and classical CNN in image classification with about 10% parameters. This study provides an effective approach for alleviating the barren plateau issue and designing multi-channel quantum convolutional neural networks for processing big classic data with fewer quantum resources.
Resource constraints and imperfect quantum operations remain major challenges for quantum machine learning (QML) on noisy intermediate-scale quantum (NISQ) devices, particularly for fault-diagnosis tasks involving imbalanced and heterogeneous data. This work proposes QTN-RQNN, a Quantum Tensor Network-Based Recurrent Quantum Neural Network that integrates tensor-network-inspired nearest-neighbor interactions with recurrent quantum processing. The model sequentially encodes input features while propagating quantum memory across recurrent steps and incorporates QSMOTE for class balancing. QTN-RQNN is evaluated on four heterogeneous fault-diagnosis datasets SPID, CWRUBD, EFDD, and SFDD and compared with QNN, QTN-QNN, and RQNN baselines through classification, ablation, multi-seed, convergence, and quantum-resource analyses. Results show dataset-dependent predictive behavior, indicating that increased quantum architectural complexity does not universally improve classification. On SPID, QTN-RQNN improves the RQNN F1-score from 0.3077 to 0.4516 and ROC-AUC from 0.4485 to 0.6970, while the remaining datasets exhibit different performance-resource trade-offs and sensitivity to recurrent timesteps and circuit depth. Across all datasets, QTN-RQNN reduces the two-qubit-gate count from 32 to 24 per sample relative to RQNN, corresponding to a 25% reduction without increasing the four-qubit requirement; the same relative reduction is obtained for QTN-QNN over QNN. These findings establish QTN-RQNN as a resource-conscious recurrent quantum architecture and highlight the need to jointly consider predictive performance and circuit-resource requirements for near-term quantum fault diagnosis.