Recent Advances for Quantum Neural Networks in Generative Learning

1 INTRODUCTION

• Classical generative models learn training-data distributions to generate new samples and have been used to address the curse of dimensionality in quantum physics.The paper attributes their classical success to deep neural networks’ ability to capture underlying data distributions.Source: §1 · p. 1
• Quantum computers are proposed as a way to address the computational overhead of classical generative models as Moore’s law approaches its limits, with QGLMs designed for NISQ machines and potential computational advantages.Reported task examples include image generation, quantum state approximation, and drug design [12].Source: §1 · p. 1
• The survey classifies quantum generative learning models as QCBMs, QGANs, QBMs, and QAEs, reviewing each family’s pioneering work, variants, classical relations, and applications.It presents this as the first review focused on quantum generative learning.Source: §1 · p. 2
• The survey proceeds from background material to QGLM literature and then to challenges and future directions.Section 2 introduces neural networks, classical generative models, quantum computing, and variational quantum algorithms; Section 3 reviews QGLMs.Source: §1 · p. 2

1.1 Related works

• Prior reviews cover fault-tolerant QML, NISQ components, and QNNs; none comprehensively reviews modern QGLM mechanisms and their relations to classical GLMs.The survey also distinguishes its scope from reviews of quantum classifiers [22] and classical GLMs applied to quantum physics.Source: §1.1 · p. 2

• To accommodate readers with diverse backgrounds, the preliminaries introduce neural networks, classical generative learning models, quantum computing, and variational quantum algorithms.The section also collects notation and frequently used acronyms in Tables 1 and 2.Source: §2 · p. 2

• Neural networks, also called artificial neural networks, process data such as images, text, and speech and are widely applied in areas including computer vision and natural language processing.The section notes their recent resurgence following the success of neural networks.Source: §2.1 · p. 2
• A perceptron is a one-neuron linear binary classifier limited to linearly separable patterns; combining perceptrons with nonlinear activations gives an MLP greater modeling capacity.MLPs have input and output layers plus hidden layers.Source: §2.1 · p. 3
• MLPs compose trainable linear transformations and nonlinear activations across layers; random initialization and backpropagation train weights against costs measuring prediction-label discrepancy.Hidden-layer activations may be ReLU, sigmoid, or tanh; output activations include softmax for classification and sigmoid for regression.Source: §2.1 · pp. 3–4
• RNNs are used for sequential data, CNNs for image analysis, and Transformers in NLP and computer vision; GPU-supported language models can exceed 100 billion parameters.The section gives GPT-3 as an example with over 100 billion parameters, trained on around 45TB of text data.Source: §2.1 · p. 4

2.2.1 Multi-layer perceptron generative network

• With fixed input and an n-dimensional output, an MLP estimates a discrete distribution; with sufficient examples and a near-minimal negative log-likelihood, its estimate approaches the target q with low error.The output probabilities correspond to possible outcomes.Source: §2.2.1 · p. 4

2.2.2 Generative adversarial network

• GANs train a generator against a discriminator in a minimax game; at Nash equilibrium, generated data is indistinguishable from target data, though training can face mode collapse and vanishing gradients.Conditional GANs can control generated outcomes, while Wasserstein GANs use a Wasserstein-derived cost to address vanishing gradients and mode collapse.Source: §2.2.2 · p. 5

2.2.3 Boltzmann machine

• Boltzmann machines model visible-state probabilities by summing over hidden states, but learning can be impractical with many hidden units; RBMs restrict within-layer connections for efficient parallel contrastive-divergence updates.The RBM restriction permits hidden and visible units to be updated in parallel using Gibbs sampling.Source: §2.2.3 · p. 5, Figure 6 · p. 5

2.2.4 (Variational) Autoencoder

• Autoencoders compress inputs into lower-dimensional latent features and reconstruct them; VAEs approximate an intractable posterior with an inference model and maximize an ELBO using reparameterized Monte Carlo gradients.Variants target posterior flexibility, disentanglement, or collapse, and VAE families are applied to images, language, and molecular design.Source: §2.2.4 · pp. 6–7

2.3.1 Quantum mechanics

• Quantum systems are represented by normalized state vectors that evolve under unitary operators, while mixed states use density matrices and measurements yield probabilistic outcomes.The expected value of an observable is given by tr(ρO).Source: §2.3.1 · p. 7

2.3.2 Quantum computing

• Qubits form quantum circuits whose gates manipulate quantum information; single-qubit rotations, phase shifts, and CNOT gates together provide a universal gate set.The Deutsch-Jozsa algorithm determines whether a promised Boolean function is constant or balanced with one oracle query.Source: §2.3.2 · pp. 7–8

2.3.3 Quantum complexity theory

• Quantum algorithms offer speedups on some tasks, but noise and qubit limits constrain implementation of algorithms such as Shor’s and Grover’s.Variational quantum algorithms are presented as an alternative design philosophy for using available NISQ devices.Source: §2.3.3 · pp. 8–9

2.4 Variational quantum algorithms

• VQAs use parameterized quantum circuits as learning models and classical optimizers to tune parameters against a cost function, forming a hybrid approach suited to near-term devices.The circuit parameters are updated to reduce the discrepancy between predictions and labels.Source: §2.4 · p. 9
• A QNN encodes classical data into a quantum state, applies a layered Ansatz, and measures observables to produce predictions.Data encoding may use basis, amplitude, or gate encoding, and the data unitary may be queried once or multiple times.Source: §2.4 · p. 9
• Ansatz design trades task-informed structure against hardware constraints: hardware-efficient circuits adapt gate connectivity to device topology, while tensor-network and neural-network Ansätze mimic classical architectures.Hamiltonian-informed Ansätze retain gates relevant to a known problem Hamiltonian; Hamiltonian-agnostic Ansätze are used when the Hamiltonian is unavailable or hardware constraints hinder informed designs.Source: §2.4 · pp. 9–10, Figure 11 · p. 10

3 QUANTUM GENERATIVE LEARNING MODELS

• The survey interprets QGLMs as quantum extensions of classical generative learning models and reviews their current progress.It introduces seminal work in each class, followed by variants and potential applications.Source: §3 · p. 10
• The survey provides tables of representative QGLM works and available source codes.Table 5 summarizes open-source codes for QGLMs.Source: §3 · p. 10, Table 5 · p. 10

3.1 Quantum circuit Born machine

• QCBMs replace the MLP generator with a QNN that learns a tunable discrete distribution under the Born rule and can be implemented on NISQ devices.The reviewed framework uses a circuit, cost function, and classical optimizer; the original implementation used KL divergence and particle swarm optimization.Source: §3.1 · pp. 10–11, Figure 12 · p. 10
• QCBM experiments on image generation, GHZ-state preparation, and thermal-state preparation show that performance depends on entangling-layer topology, with deeper circuits potentially helping harder tasks.Trained QCBMs recovered compact GHZ preparation protocols on trapped-ion computers.Source: §3.1 · p. 11

3.1.1 Variants

• QCBM variants alter Ansatz structure, optimization, or resources; examples include sparse connectivity, RY-only gates, ancillary qubits, and tensor-network Ansätze.Reported comparisons include local minima for some entanglement configurations and fewer qubits and parameters for a tensor-network approach with comparable generative capability.Source: §3.1.1 · pp. 11–12
• RY gates can enable fast training but induce circuit rigidity, whereas arbitrary rotation gates can benefit deeper circuits by improving loss-landscape connectivity.Source: §3.1.1 · p. 12
• QCBM studies use KL divergence, MMD, and alternatives including Stein discrepancy, Sinkhorn divergence, and Jensen-Shannon divergence; Sinkhorn divergence was reported superior to MMD on certain tasks.MMD can be more efficiently applied in large-scale systems when KL divergence is inaccessible.Source: §3.1.1 · p. 12
• Gradient-based optimizers suit large QNNs, while gradient-free methods suit smaller models or highly non-smooth costs; CMA-ES was reported more noise-vulnerable and slower to converge than Adam.The survey also describes derivative-free and gradient-based optimizer families, including PSO, Bayesian optimization, Adam, and L-BFGS-B.Source: §3.1.1 · pp. 12–13

3.1.2 Applications

• QCBMs have been applied to financial distributions, synthetic data, and joint distributions, with some studies reporting better or equivalent performance to classical RBMs, especially at larger scales.A reported stock-market copula model used up to 8 qubits; state preparation is possible under conditions such as integrability.Source: §3.1.2 · p. 13

3.2 Quantum generative adversarial network

• QGANs use generator–discriminator minimax learning, with quantum neural networks available to implement either player and targets that may be discrete or continuous.The survey distinguishes four configurations according to task type and whether the generator and discriminator are quantum or classical.Source: §3.2 · p. 13
• QT-QGQD alternates discriminator and generator updates; its Nash equilibrium cost reaches zero only when the generated state equals the target state.For a fixed generator, the discriminator optimum is the Helstrom measurement; the equilibrium condition is stated in Ref..Source: §3.2 · p. 13
• For continuous classical distributions, CT-QGCD is described as potentially advantageous because quantum processors represent N-dimensional vectors with log N qubits.The passage contrasts claimed optimization runtime O(poly(logN)) with O(poly(N)) for classical GANs.Source: §3.2 · pp. 13–14

3.2.1 Variants

• QT-QGQD studies include conditional QuGANs, quantum-state estimation, superconducting-device demonstrations, and alternative cost functions for robustness or fewer parameters.Reported examples include learning multi-qubit states and the XOR gate [180], a single-qubit state, and quantum Wasserstein and fidelity costs [143] [138].Source: §3.2.1 · p. 14
• Other variants report quadratic cost-function sample-complexity reduction, comparable semi-supervised performance with fewer parameters, and continuous-distribution modeling.The survey also describes a proposal intended to avoid vanishing gradients for discrete distributions and a quantum generator reported to reduce parameters in drug discovery.Source: §3.2.1 · pp. 14–15

3.2.2 Applications

• QGAN applications include quantum-state approximation, financial pricing and volatility modeling, handwritten-digit generation, and drug discovery.Patch-QGAN and batch-QGAN were experimentally demonstrated to generate real-world handwritten digits on superconducting quantum machines.Source: §3.2.2 · p. 15

3.3 Quantum Boltzmann machine

• A QBM models discrete target distributions with a Hamiltonian-based Boltzmann distribution, and trains parameters by minimizing negative log-likelihood.The semi-restricted variant disallows internal connections between hidden units.Source: §3.3 · p. 15
• Exact negative-phase sampling is NP-hard, making its calculation intractable for large-scale problems; a Golden–Thompson upper bound offers an alternative cost.That bound permits efficient updates of {b, w}, but does not train the transverse field Γ, which must be preset.Source: §3.3 · p. 15
• Reported simulations find improved trainability over classical models for Bernoulli mixtures and lower cost for a joint-distribution supervised task.A D-Wave 2X implementation with up to 940 qubits demonstrated feasibility for unsupervised generative learning.Source: §3.3 · pp. 15–16

3.3.1 Variants

• QBM variants use bound-based, state-based, or circuit-based training, including POVM measurements, relative-entropy objectives, and variational quantum imaginary-time evolution.POVM-based training allows transverse-field training; VarQITE prepares approximate Gibbs states and has been demonstrated on quantum hardware.Source: §3.3.1 · pp. 16–17

3.3.2 Applications

• QBM applications include maze-traversal reinforcement learning, where simulations reported the best performance for QBM, and language processing using a Fock-space representation.In the maze task, DBM outperformed RBM as training samples increased.Source: §3.3.2 · p. 17

3.4 Quantum autoencoder

• A QAE encodes an input state into subsystem A while measuring subsystem B, then decodes the compressed state with a reference state to reconstruct the input.Its cost is the weighted average infidelity between input and output states; zero cost corresponds to exact recovery.Source: §3.4 · p. 17

3.4.1 Variants

• QAE studies address compression, reconstruction, preparation, and denoising; proposed improvements include hybrid schemes and alternative encoder–decoder structures.Deep parameterized circuits may still encounter barren plateaus or narrow gorges, while lower-depth circuits may have local minima.Source: §3.4.1 · pp. 17–18
• For compression, EF-QAE uses local cost functions and feature-vector encoding, with simulations reporting higher fidelity than standard QAEs.The comparison covered transverse-field Ising ground states and classical handwritten digits; optimal compression is impossible when input-state rank exceeds compressed-space dimension.Source: §3.4.1 · pp. 17–18
• Photonic QAE experiments reported low error rates when compressing qutrits to qubits and robustness to channel drift during optimization.Other reconstruction proposals use amplitude-damping channels to increase output-state rank or an adiabatic model to handle large-scale problems.Source: §3.4.1 · p. 18
• QAEs have also been proposed for quantum-state preparation, denoising, error mitigation, and classical tasks including handwritten-digit reconstruction and clustering.A quantum variational autoencoder appends a QBM to represent a classical variational autoencoder’s discrete latent space and was reported to attain speedups and require less storage than classical methods.Source: §3.4.1 · p. 18

4 CONCLUSIONS AND DISCUSSIONS

• The survey reviews QGLMs’ potential benefits for conventional machine learning and quantum physics, while noting that the field remains in its infancy.It discusses these prospects in terms of computational efficiency and model expressivity.Source: §4 · p. 19

4.1 Challenges

• Read-in and read-out bottlenecks make current QGLMs inferior for high-dimensional continuous distributions, motivating more efficient data-loading and information-extraction strategies.Amplitude encoding can require loading classical examples into quantum states, while extracting image features may require quantum state tomography with complexity scaling with feature dimension.Source: §4.1 · p. 19
• QGLM learnability remains poorly understood: generalization definitions vary by model, and non-convex losses complicate trainability and global-minimum analysis.Generalization evidence covers only specified settings, while strategies for barren plateaus beyond some QBMs and unitary QNNs, and for reaching optimal or near-optimal results, remain unclear.Source: §4.1 · p. 19

4.2 Prospects

• Noise can degrade QGLM performance or diverge optimization, while shallow hardware-efficient circuits trade reduced noise effects for diminished model power and computational advantages.Future directions include error mitigation and problem-specific, hardware-oriented Ansätze; their effectiveness has been validated in quantum discriminative learning, but little is known about QGLMs.Source: §4.1 · p. 19, §4.2 · p. 20
• For read-in, the survey calls for NISQ-suitable amplitude encoding; for read-out, it proposes adapting fault-tolerant methods or using random measurements such as shadow tomography.Zhang et al. [278] describe state-preparation depth bounds using exponentially many ancillary qubits; extending such results to tolerable-error NISQ settings remains an open question.Source: §4.2 · pp. 19–20
• Learnability research could apply statistical tools to QGLMs, combine architecture search with prior knowledge, and test whether overparameterized models retain desirable training behavior.Underparameterized designs aim to keep the hypothesis space relatively small while covering the optimal solution; whether discriminative-model initialization strategies transfer to overparameterized QGLMs is unclear.Source: §4.2 · p. 20