quant-phFeb 26, 2026

Quantum Deep Learning: A Comprehensive Review

Authors: Yanjun Ji, Zhao-Yun Chen, Marco Roth, David A. Kreplin, Christian Schiffer, Martin King, Oliver Anton, M. Sahnawaz Alam, 

Organizations: Institute for Quantum Computing Analytics (PGI-12), Forschungszentrum Jülich, 52425 Jülich, Germany · Institute of Artificial Intelligence, Hefei Comprehensive National Science Center, Hefei, Anhui, 230088, P. R. China · Fraunhofer Institute for Manufacturing Engineering and Automation (IPA), Nobelstrasse 12, Stuttgart 70569, Germany · Heilbronn University of Applied Sciences, Max-Planck-Str. 39, 74081 Heilbronn, Germany · 

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Quantum deep learning (QDL) explores the use of both quantum and quantum-inspired resources to determine when deep learning's core capabilities, such as expressivity, generalization, and scalability, can be enhanced based on specific resource constraints. Distinct from broader quantum machine learning, QDL emphasizes compositional depth at the pipeline level and the integration of quantum or quantum-inspired components within end-to-end workflows. This review provides an operational definition of QDL and introduces a taxonomy comprising four primary paradigms: hybrid quantum-classical models, quantum deep neural networks, quantum algorithms for deep learning primitives, and quantum-inspired classical algorithms. Theoretical principles are connected to advanced architectures, software toolchains, and experimental demonstrations across superconducting, trapped-ion, photonic, semiconductor spin, and neutral-atom systems, as well as quantum annealers. Claims of quantum advantage are critically assessed by distinguishing provable complexity-theoretic separations from empirical observations. The analysis characterizes trade-offs between model expressivity, trainability, and classical simulability, while systematically detailing the bottlenecks imposed by optimization landscapes, input-output access models, and hardware constraints. Applications are surveyed in domains encompassing image classification, natural language processing, scientific discovery, quantum data processing, and quantum optimal control, underscoring fair benchmarking against optimized classical counterparts and a comprehensive assessment of resource requirements. This review serves as a tutorial entry point for graduate students while guiding readers to specialized literature. It concludes with a verification-aware roadmap to transition QDL from near-term demonstrations to scalable and fault-tolerant implementations.

Figures & tables

Figure 1: Four paradigms of quantum deep learning (QDL), ordered left to right by increasing centrality of the quantum component to the end-to-end model. (a) Quantum-inspired algorithms: purely classical models in which the structure or training algorithms are motivated by quantum information theory. Quantum content is expressed through the model class, not through hardware-executed quantum states. (b) Hybrid quantum-classical models: a parameterized quantum circuit (PQC) with trainable gate parameters θ\bm{\theta} is embedded as a differentiable module within a classical pipeline. The cycle denotes the classical outer optimization loop that updates θ\bm{\theta} from measurement outcomes by minimizing a loss function L(θ)\mathcal{L}(\bm{\theta}) . (c) Quantum algorithms for deep learning (DL) primitives: the quantum processing unit (QPU) acts as a specialized coprocessor to accelerate core subroutines, e.g., linear-system solvers such as HHL and the quantum singular value transformation (QSVT), within an overarching classical model under explicitly stated access and resource assumptions. (d) Quantum deep neural networks: deep, hierarchical quantum circuits designed for end-to-end quantum information processing, e.g., feature extraction directly from quantum data. Rx(θi)R_{x}(\theta_{i}) and Rz(θi)R_{z}(\theta_{i}) : single-qubit rotation gates; XX : Pauli- XX gate; HH : Hadamard gate.
Figure 2: The hybrid quantum-classical learning loop. Classical data xx are encoded into an input state ρin(x)\rho_{\mathrm{in}}(x) via the data-encoding unitary S(x)S(x) . A parameterized quantum circuit (PQC) U(θ)U(\bm{\theta}) is executed, measurements of an observable yield finite-shot statistics used to estimate a loss function (e.g., expectation value f(x;θ)f(x;\bm{\theta}) ), and a classical optimizer updates θ\bm{\theta} by minimizing an objective built from these estimates. The loop repeats until convergence to an optimal parameter set θ∗\bm{\theta}^{*} .
Figure 3: Taxonomy of hybrid quantum-classical architectures across four paradigms. (a) Placement and composition. A classical encoder compresses the input xx to a latent code zz , which is embedded via S(z)S(z) ; this encoding of zz rather than the raw data xx directly alters the induced hypothesis class. An interleaved-exchange variant (lower) routes xx through alternating quantum modules Q1,Q2,…Q_{1},Q_{2},\ldots and classical layers where each QiQ_{i} subsumes the full encode-process-measure pipeline S(⋅)→U(θi)→MS(\cdot)\to U(\bm{\theta}_{i})\to\mathcal{M} of the upper subpanel. (b) Hierarchical and geometric structure. Quantum convolutional neural network (QCNN, upper): alternating trainable-unitary convolutional and measurement-based pooling layers act on ρin\rho_{\mathrm{in}} , followed by a fully connected final mixing layer; Graph models (lower): node and edge unitaries UvU_{v} , UeU_{e} implement permutation-equivariant message passing on G=(V,E)G{=}(V,E) . (c) Interfaces and encoding patterns. Quanvolution (upper): local PQCs process image patches xix_{i} and output a spatial feature map for classical aggregation. Data reuploading (lower left): interleaving S(x)S(x) with layers U(θl)U(\bm{\theta}_{l}) implements a Fourier-feature expansion. Independent and multi-channel (lower right): independent channels concatenate per-channel measurement outcomes classically, whereas in multi-channel, joint encoding applies coherent premeasurement mixing, enabling coherent cross-channel mixing in the quantum feature map prior to measurement. (d) Objective-driven paradigms. Training loop (upper): a classical backbone, quantum primitive U(θ)U(\bm{\theta}) , and task objective jointly update θ\bm{\theta} ; instantiations include quantum generative adversarial networks (QGAN) discriminators, Boltzmann models, and reinforcement learning (RL) value functions. Deployment loop (lower): U(θ∗)U(\bm{\theta}^{*}) operates at trained and fixed parameters under a sampling bottleneck on quantum devices, producing scores, samples, or actions.
ParadigmQuantum-module roleInductive biasKey refs.
Placement and composition (graph location)
PQC headterminal readout / classifierencoding-induced hypothesis restriction; restricted readout[ 237 , 719 ]
Transfer hybridquantum head on frozen classical featuresrepresentation bottleneck; low-dimensional embedding[ 546 , 162 ]
Interleaved exchangemultiple quantum blocks separated by classical processingnon-sequential composition; measurement-mediated feature exchange across modules[ 500 , 525 , 299 , 398 ]
Hierarchical and geometric structure (symmetry + locality priors)
QCNNhierarchical poolingmultiscale locality via pooling; parameter tying[ 179 , 331 ]
Table 1: Representative taxonomy of hybrid quantum-classical models. Architectures are classified by the computational role of the quantum circuit and the primary structural constraint that shapes the hypothesis class and associated estimator family. Figure 3 depicts canonical core motifs for each paradigm; the table lists representative instantiations and closely related variants. Abbreviations: PQC, parameterized quantum circuit; QCNN, quantum convolutional neural network; QNN, quantum neural network; RL, reinforcement learning; QGAN, quantum generative adversarial network.
Figure 4: Fundamental tensor network (TN) architectures. Shown are matrix product state (MPS; also known as tensor train), projected entangled pair state (PEPS), tree tensor network (TTN), and the multiscale entanglement renormalization ansatz (MERA). The top row illustrates increasing tensor order (scalar, vector, matrix, tensor) as a visual legend for node types. Adapted from Berezutskii et al. [83] .
Figure 5: Resource complexities under an explicit access model A\mathcal{A} . (a) three physical access regimes, including single-copy prepare-and-measure; collective access, where kk copies of ∣ψ⟩|\psi\rangle per query are processed jointly via quantum memory ( kk denotes the number of copies jointly processed per example instance); and coherent oracle or QRAM-style state access. (b) A\mathcal{A} constrains hypothesis-class complexity CH\mathcal{C_{H}} , which in turn mediates trade-offs among three consumable resource axes: example complexity CN\mathcal{C_{N}} , measurement complexity CS\mathcal{C_{S}} (total circuit repetitions, i.e., shots ×\times settings, typically incurred per training example for a fixed estimator precision), and computational complexity CT\mathcal{C_{T}} . The number of total state preparations per objective evaluation is CN×CS\mathcal{C_{N}}\times\mathcal{C_{S}} and scales as k\penalty CN×CSk\penalty\ \mathcal{C_{N}}\times\mathcal{C_{S}} under collective access.
Figure 6: Five-layer implementation stack for QDL, illustrating top-down requirements flow and bottom-up constraint propagation.
ModalityPrimitiveConnectivityQDL bottleneck
Superconductingdigital gatessparse 2D (routing)routing overhead, drift/crosstalk
Trapped ionsdigital gatesall-to-all (in-modular); modular linkswall-clock throughput
Neutral atomsdigital gates + analogreconfigurable between shotsentangling error, atom loss
Photonicsoptical sampling; CVinterferometric meshes; delay networksoptical loss, calibration
Annealersanalog samplingfixed graph (minor embedding)embedding, non-equilibrium bias
Semiconductor spindigital gateslocal couplingvariability, coupling range
Table 2: Modality-level hardware characteristics most relevant to QDL feasibility and codesign. Connectivity refers to native two-qubit (or native entangling) interaction structure; QDL bottleneck denotes a typical dominant limitation for end-to-end hybrid training at current scales. CV: continuous-variable. For annealers, minor embedding refers to mapping a logical problem graph onto the fixed hardware graph.
SoftwareClassical SimulationHybrid WorkflowQPU Integration
GPU-enabledHPC-enabledTNPyTorchJAXKerasCircuit-basedPhotonics
PennyLane✓✓✓✓✓✓✓✓
TensorFlow Quantum✓✓×\times×\times×\times✓✓×\times
Qiskit Machine Learning✓✓✓✓×\times×\times✓×\times
CUDA-Q✓✓✓✓×\times×\times✓×\times
sQULearn✓∘\circ×\times✓×\times×\times✓×\times
Table 3: Comparison of ten mainstream QDL software frameworks. The evaluation spans classical simulation capabilities, integration with classical libraries for hybrid workflows, and support for diverse quantum hardware paradigms. GPU-enabled: officially supported GPU-accelerated simulation backend; HPC-enabled: documented multi-node and distributed execution; TN: documented tensor network (TN) approximate simulators integrated in the stack.
Figure 7: Taxonomy of QDL application regimes by data interface. Classical data pipelines embed classical features into quantum circuits and are typically bottlenecked by state-preparation overheads. Scientific discovery integrates quantum modules into label-expensive workflows with costly forward models for both classical and quantum data, where sample efficiency is paramount. Quantum data pipelines operate directly on quantum-native states or processes under an explicit access model, with bottlenecks set by the readout interface and the availability of coherent quantum memory. The lower panel summarizes the four-pillar evaluation protocol used in Sec. V.2 .
PillarReporting axesPrimary failure modes
1. Empirical performance Does it improve quality under matched evaluation?– Task-specific metric and calibration – Learning curves versus data and compute – Robustness across seeds and splits – Sample-efficiency diagnostics– Under-tuned or outdated baselines – Unmatched hyperparameter optimization – Hidden pre or postprocessing – Selective reporting of best runs
2. Resource complexity Is it more efficient under end-to-end accounting?– Total shots and QPU calls – Wall-clock breakdown and time-to-target – Energy and monetary cost boundary – Compiled depth and two-qubit counts – Classical coprocessing and communication– Reporting circuit metrics without wall-clock cost – Incommensurate access or readout interfaces or training interfaces (e.g., classical backpropagation) – State-preparation and encoding omitted – Missing classical surrogate checks
3. Practical feasibility Can it be trained and executed reliably on quantum hardware?– Training stability and failure rate – Verification method and its cost – Inference throughput and latency – Compilation overheads– Optimizer stagnation and seed sensitivity – Shot-noise dominated updates – Unsupported intractability
4. Domain alignment Is the quantum component structurally matched to the task?– Architecture class and symmetry constraints – Modality and interface realism – Structure-sensitive ablations and controls – Inductive-bias diagnostics– Access-model mismatch to the deployment setting – Baselines that ignore the same structure – Claims of inductive bias without controls
Table 4: A four-pillar protocol for rigorous quantum-classical comparison in finite-size empirical studies. Claims are evaluated under an explicit access and readout interface and an end-to-end resource contract R\mathcal{R} (Definition 3). Mechanisms underlying limitations are treated in Sec. III.7 .
Figure 8: Strategic roadmap for QDL, illustrating the coupled evolution of four dimensions: hardware scale, error management, algorithmic sophistication, and verification rigor. We distinguish three capability regimes by heuristic operation budgets and error-handling assumptions [ 228 , 655 ] . Here, an “elementary operation” means one primitive physical action in a single compiled circuit execution (one QPU call), counting each native single-qubit gate, two-qubit gate, and single-qubit measurement once (summed over all qubits). Under this convention: (i) the near-term NISQ regime spans ∼104\sim 10^{4} – 10610^{6} elementary operations per circuit execution, dominated by hybrid training and error mitigation; (ii) an early fault-tolerant regime spans ∼106\sim 10^{6} – 10910^{9} , enabling deeper circuits and explicitly QEC-aware training protocols; and (iii) an application-scale fault-tolerant regime spans ∼109\sim 10^{9} – 101210^{12} , where fully fault-tolerant quantum deep architectures and resource-intensive algorithmic subroutines become feasible.
More in this topic

Quantum Machine Learning

Quantum machine learning uses quantum circuits and quantum data to build learning algorithms and study their computational and statistical limits.

See all 256 papers

Explore similar work

CardsList
Jun 30, 2026

Beyond the Expressivity-Trainability Paradox: A Dynamical Lie Algebra Perspective on Navigating Barren Plateaus in Quantum Machine Learning

As Quantum Machine Learning (QML) transitions toward practical implementation, the field faces a critical architectural bottleneck that challenges the fundamental assumptions of classical statistical learning theory. In classical deep learning, increasing model capacity typically risks overfitting. However, this study advances a counter-intuitive paradigm: unstructured contemporary QML architectures suffer from a profound state of quantum underfitting, driven by the "expressivity-trainability paradox." We demonstrate that the vast Hilbert space capacity of Parameterized Quantum Circuits (PQCs)-traditionally chased as the source of quantum advantage is the direct mathematical cause of Barren Plateaus (BPs), where gradient landscapes become exponentially flat. By synthesizing recent breakthroughs in Dynamical Lie Algebras (DLAs) and Geometric QML, we establish a comprehensive framework linking the algebraic dimension of circuit generators to their optimization dynamics. Furthermore, we empirically validate this framework on a non-linear binary classification task, illuminating a uniquely quantum manifestation of the bias-variance tradeoff: while unstructured architectures achieve near-perfect training accuracy via unscalable parameterization (quantum overfitting), embedding group-theoretic geometric priors acts as a structural regularizer. By restricting the DLA growth to a polynomial regime, our symmetry-preserving approach sacrifices raw memorization capacity to guarantee scalable, gradient-rich training landscapes, offering a robust roadmap for "Trainability-by-Design" in scalable quantum neural networks.

Kung-Ming Lan, Edward Huang

Department of Industrial Engineering and Enterprise Information, Tunghai University, Taichung 40704, Taiwan · Department of Industrial and Systems Engineering, 3306 Shelby Center, Auburn University, Auburn, AL, USA

Quantum Machine LearningBarren PlateausVariational Quantum CircuitsQuantum Neural Networks
AskPDF
Aug 2, 2026

Hybrid Quantum Neural Networks: Theory, Implementations, and Applications

Artificial intelligence has been transformed by deep neural networks, yet the search for new learning architectures continues. Quantum machine learning offers one such direction, and hybrid quantum neural networks, which combine classical neural-network components with quantum information processing units, have emerged as a practical framework for near-term quantum technologies. However, the rapid development of the field across diverse architectures, benchmarks and hardware assumptions makes it difficult to assess the utility of various proposals, identify where genuine advantages may arise, and determine how practitioners can use these models. While recent benchmarks caution that such gains have not yet been demonstrated at scale, theoretical work has identified tasks on which quantum models hold provable advantages, and hybrid approaches have delivered promising results on practical problems using deliberately compact quantum components and substantially fewer trainable parameters. Here, we review hybrid quantum neural networks for the machine-learning and quantum-machine-learning communities. We summarize their main theoretical and methodological foundations, survey some of the most promising architectures developed so far, and examine their implementation challenges and reported performance. By consolidating these perspectives, this review provides a structured view of the state of the field and helps identify promising paths for future research and application-driven development.

Léo Monbroussou, Maniraman Periyasamy, Viacheslav Kuzmin +4

Terra Quantum AG, Kornhausstrasse 25, St. Gallen, 9000, Switzerland · Département de Physique Appliquée, Université de Genève, 24 Quai Ernest-Ansermet, Geneva, 1211, Switzerland

Hybrid Quantum-Classical MLVariational Quantum CircuitsQuantum Neural Networks
AskPDF
Jun 12, 2026

Quantum Machine Learning for Industrial Applications

Recent advances in Machine Learning have transformed numerous industrial sectors, yet classical paradigms face fundamental limitations: rapidly growing data volumes, rising computational costs, significant energy consumption, and the physical scaling limits of conventional hardware architectures. Quantum computing has emerged as a promising computational paradigm to address these challenges, giving rise to the field of Quantum Machine Learning (QML). In this thesis, the theoretical foundations of QML are investigated, with a focus on near-term and future practical applications. Three central challenges are addressed: the trainability of variational quantum circuits, their expressivity, and their resistance to efficient classical simulation. The trainability of Hamming-weight preserving variational quantum circuits is first studied, and theoretical guarantees are established that resolve an open conjecture on the absence of barren plateaus for this circuit family. Subspace-preserving QML algorithms are then introduced, including photonic circuits and quantum convolutional neural networks, and are designed to mimic classical ML subroutines while offering polynomial quantum advantage. Finally, variational quantum circuits are analyzed as quantum Fourier models, and a framework is derived to jointly characterize expressivity and trainability, from which conditions are obtained under which quantum models provably separate from their classical counterparts. These contributions are intended to advance the theoretical roadmap for harnessing near-term and future quantum technologies in real-world applications.

Léo Monbroussou

Sorbonne Université, Ecole Doctorale Informatique, Télécommunications et Electronique (n° 572), Computer Science

Quantum Convolutional Neural NetworksQuantum Machine LearningBarren PlateausVariational Quantum CircuitsQuantum Neural Networks
AskPDF