Quantum Deep Learning: A Comprehensive Review

TL;DR

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, and a verification-aware roadmap to transition QDL from near-term demonstrations to scalable and fault-tolerant implementations.

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 (8)

• 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 $\boldsymbol{\theta}$ is embedded as a differentiable module within a classical pipeline. The cycle denotes the classical outer optimization loop that updates $\boldsymbol{\theta}$ from measurement outcomes by minimizing a loss function $\mathcal{L}(\boldsymbol{\theta})$. (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. $R_x(\theta_i)$ and $R_z(\theta_i)$: single-qubit rotation gates; $X$: Pauli-$X$ gate; $H$: Hadamard gate.
• Figure 2: The hybrid quantum-classical learning loop. Classical data $x$ are encoded into an input state $\rho_{\mathrm{in}}(x)$ via the data-encoding unitary $S(x)$. A parameterized quantum circuit (PQC) $U(\boldsymbol{\theta})$ is executed, measurements of an observable yield finite-shot statistics used to estimate a loss function (e.g., expectation value $f(x;\boldsymbol{\theta})$), and a classical optimizer updates $\boldsymbol{\theta}$ by minimizing an objective built from these estimates. The loop repeats until convergence to an optimal parameter set $\boldsymbol{\theta}^*$.
• Figure 3: Taxonomy of hybrid quantum-classical architectures across four paradigms. (a) Placement and composition. A classical encoder compresses the input $x$ to a latent code $z$, which is embedded via $S(z)$; this encoding of $z$ rather than the raw data $x$ directly alters the induced hypothesis class. An interleaved-exchange variant (lower) routes $x$ through alternating quantum modules $Q_1, Q_2, \ldots$ and classical layers where each $Q_i$ subsumes the full encode-process-measure pipeline $S(\cdot) \to U(\boldsymbol{\theta}_i) \to \mathcal{M}$ 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 $\rho_{\mathrm{in}}$, followed by a fully connected final mixing layer; Graph models (lower): node and edge unitaries $U_v$, $U_e$ implement permutation-equivariant message passing on $G{=}(V,E)$. (c) Interfaces and encoding patterns. Quanvolution (upper): local PQCs process image patches $x_i$ and output a spatial feature map for classical aggregation. Data reuploading (lower left): interleaving $S(x)$ with layers $U(\boldsymbol{\theta}_l)$ 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 $U(\boldsymbol{\theta})$, and task objective jointly update $\boldsymbol{\theta}$; instantiations include quantum generative adversarial networks (QGAN) discriminators, Boltzmann models, and reinforcement learning (RL) value functions. Deployment loop (lower): $U(\boldsymbol{\theta}^{*})$ operates at trained and fixed parameters under a sampling bottleneck on quantum devices, producing scores, samples, or actions.
• 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 berezutskii2025tensor.
• Figure 5: Resource complexities under an explicit access model $\mathcal{A}$. (a) three physical access regimes, including single-copy prepare-and-measure; collective access, where $k$ copies of $|\psi\rangle$ per query are processed jointly via quantum memory ($k$ denotes the number of copies jointly processed per example instance); and coherent oracle or QRAM-style state access. (b) $\mathcal{A}$ constrains hypothesis-class complexity $\mathcal{C_H}$, which in turn mediates trade-offs among three consumable resource axes: example complexity $\mathcal{C_N}$, measurement complexity $\mathcal{C_S}$ (total circuit repetitions, i.e., shots $\times$ settings, typically incurred per training example for a fixed estimator precision), and computational complexity $\mathcal{C_T}$. The number of total state preparations per objective evaluation is $\mathcal{C_N} \times \mathcal{C_S}$ and scales as $k~\mathcal{C_N} \times \mathcal{C_S}$ under collective access.