Hybrid Quantum-Trained Neural Networks

• Hybrid QT-NNs are systems that integrate parameterized quantum circuits with classical architectures to jointly optimize complex predictive tasks.
• They employ diverse strategies such as intermediate quantum layers, parallel hybridization, and quantum-to-classical mappings to enhance expressivity and training efficiency.
• Empirical evaluations demonstrate substantial parameter compression and competitive performance across classification, regression, and scientific modeling benchmarks.

Hybrid Quantum-Trained Neural Networks (QT-NNs) are hybrid quantum-classical learning systems in which parameterized quantum circuits, quantum neural networks, or other quantum modules are coupled to classical neural architectures and optimized on a shared task objective. In the literature, this umbrella includes end-to-end tensor-network/variational-circuit classifiers, hybrid networks with an intermediate quantum layer, parallel classical–quantum models, and “Quantum-Train” frameworks in which a quantum module generates parameters for a purely classical network used at inference time (Chen et al., 2021, Liu et al., 2024, Lin et al., 2024). The common design premise is that quantum and classical components should not be treated as isolated subsystems: their interaction determines expressivity, trainability, compression behavior, and deployment constraints (Kashif et al., 25 May 2026).

1. Terminological scope and model families

The designation “QT-NN” does not refer to a single canonical architecture. Closely related papers use labels such as hybrid quantum-classical neural network, hybrid quantum neural network (HQNN), Quantum-Train (QT), quantum-trained convolutional neural network (QT-CNN), and hybrid quantum physics-informed neural network (HQ-PINN). Across these variants, two broad formulations recur.

The first formulation places the quantum module directly in the prediction path. Examples include the end-to-end trainable MPS–VQC classifier, where a matrix product state compresses an input and a variational quantum circuit performs the final classification (Chen et al., 2020); the “sandwiched” architecture for continuous-variable prediction, where a QNN layer sits between classical input and output layers and is exposed through qml.qnn.KerasLayer (Jain et al., 2022); the feedforward binary classifier in which each neuron is itself a variational quantum circuit (Arthur et al., 2022); the Parallel Hybrid Network, which processes the same input through an MLP and a VQC and linearly combines their outputs (Kordzanganeh et al., 2023); and the non-sequential TunnElQNN, which alternates classical layers with a quantum layer and uses the Tunnelling Diode Activation Function in the classical components (Abbas, 2 May 2025).

The second formulation uses the quantum system primarily during training, not during inference. In the original Quantum-Train framework, a QNN with $N=\lceil \log_2 M\rceil$ qubits is measured, and a classical mapping model converts basis-state information and measurement probabilities into the $M$ weights of a classical neural network (Liu et al., 2024). QT-CNN adopts the same principle for deepfake-audio detection: QNN blocks generate compressed quantum states, a classical mapping network translates measurement statistics into CNN parameters, and the trained CNN is later deployed on classical hardware only (Lin et al., 2024). Tensor-network mapping and photonic-QT variants preserve this training-time quantum role while replacing the mapping model or the quantum substrate (Liu et al., 2024, Chen et al., 13 May 2025).

A useful generic abstraction for direct-processing hybrids is

$$y = g_{\phi}\!\left(f_{\theta}\!\left(h_{\phi}(x)\right)\right),$$

where $h_{\phi}$ is a classical feature map, $f_{\theta}$ is a PQC, and $g_{\phi}$ is a classical output map (Kashif et al., 25 May 2026). QT-style hypernetwork models instead separate data processing from quantum computation: classical data enter a classical network whose weights are synthesized by a quantum-to-classical pipeline (Liu et al., 2024).

2. Architectural motifs and representation strategies

One major motif is classical compression followed by quantum classification. In the MPS–VQC classifier, an $N$-dimensional normalized input $\mathbf{x}=(x_1,\ldots,x_N)$ is mapped componentwise as

$$x_i \mapsto \begin{bmatrix} \cos \left( \frac{\pi}{2} x_i \right) \ \sin \left( \frac{\pi}{2} x_i \right) \end{bmatrix},$$

and then lifted to the product state

$$|\Phi(\mathbf{x})\rangle = \bigotimes_{i=1}^N \begin{bmatrix} \cos \left( \frac{\pi}{2} x_i \right) \ \sin \left( \frac{\pi}{2} x_i \right) \end{bmatrix}.$$

A matrix product state contracts this feature tensor to a low-dimensional vector, which is then encoded into a VQC using rotations $M$0 and $M$1, together with Hadamard gates, CNOT entanglers, and trainable single-qubit $M$2 gates (Chen et al., 2020).

A second motif is the intermediate quantum layer used as a nonlinear transformation inside a classical pipeline. In the continuous-variable predictor, classical features are encoded with PennyLane’s AngleEmbedding, processed by StronglyEntanglingLayers, and measured as expectation values that feed a dense output layer (Jain et al., 2022). In seismic inversion, the HQ-PINN encoder uses amplitude embedding, followed by “Basic entangler layers” composed of $M$3, $M$4, or $M$5 rotations and closed-chain CNOTs; the resulting Pauli-$M$6 expectation values are passed to a classical fully connected layer, and a physics-based decoder reconstructs seismic data from the predicted elastic parameters (Vashisth et al., 6 Mar 2025).

A third motif is parallel or alternating hybridization rather than strict serial composition. In the Parallel Hybrid Network,

$M$7

with $M$8 produced by an MLP arm, $M$9 by a VQC arm, and $$y = g_{\phi}\!\left(f_{\theta}\!\left(h_{\phi}(x)\right)\right),$$0 trainable combination weights. The quantum arm is interpreted as learning a truncated Fourier series, while the classical arm contributes non-harmonic corrections (Kordzanganeh et al., 2023). TunnElQNN likewise relies on non-sequential hybridization, but its classical nonlinearity is the Tunnelling Diode Activation Function, whose negative-differential-resistance region is used before and after a 2-qubit AngleEmbedding plus BasicEntanglerLayer block (Abbas, 2 May 2025).

QT-style models replace quantum feature extraction with quantum parameter generation. In Quantum-Train, the QNN prepares a state, measurements produce probabilities $$y = g_{\phi}\!\left(f_{\theta}\!\left(h_{\phi}(x)\right)\right),$$1, and a classical mapping model $$y = g_{\phi}\!\left(f_{\theta}\!\left(h_{\phi}(x)\right)\right),$$2 receives the binary encoding of the index $$y = g_{\phi}\!\left(f_{\theta}\!\left(h_{\phi}(x)\right)\right),$$3 concatenated with the corresponding probability to output a classical parameter $$y = g_{\phi}\!\left(f_{\theta}\!\left(h_{\phi}(x)\right)\right),$$4 (Liu et al., 2024). QT-CNN refines this idea by mapping each parameter through a small classical network $$y = g_{\phi}\!\left(f_{\theta}\!\left(h_{\phi}(x)\right)\right),$$5 fed with basis bits and the measurement probability $$y = g_{\phi}\!\left(f_{\theta}\!\left(h_{\phi}(x)\right)\right),$$6, explicitly allowing generated CNN parameters to take both positive and negative values (Lin et al., 2024). Tensor-network mapping models further replace the MLP mapper with an MPS factorization, using

$$y = g_{\phi}\!\left(f_{\theta}\!\left(h_{\phi}(x)\right)\right),$$7

where $$y = g_{\phi}\!\left(f_{\theta}\!\left(h_{\phi}(x)\right)\right),$$8 is expressed as a matrix product state and $$y = g_{\phi}\!\left(f_{\theta}\!\left(h_{\phi}(x)\right)\right),$$9 is a tensorized feature map of basis bits and probabilities (Liu et al., 2024).

3. Training procedures and optimization regimes

The dominant training regime is end-to-end gradient-based optimization. In the MPS–VQC classifier, gradients flow through both tensor-network contraction and quantum simulation so that all MPS tensors and VQC gates are updated simultaneously by Adam or RMSProp on a classification loss (Chen et al., 2020). Hybrid QNN regression uses stochastic gradient descent with mean squared error, while PennyLane automatic differentiation computes quantum gradients through the parameter-shift rule or finite differences (Jain et al., 2022). VQNet generalized this style of hybrid backpropagation earlier by introducing symbolic variables, placeholders, classical operators, and quantum operators qop and qop_pmeasure, thereby treating VQCs as trainable operators embedded in a computational graph (Chen et al., 2019).

For parameterized quantum circuits, the standard gradient primitive remains the parameter-shift rule,

$h_{\phi}$0

used explicitly in hybrid classifiers and QT-style training pipelines (Arthur et al., 2022, Liu et al., 2024). In molecular QSPR modeling, hybrid quantum neural networks attach a variational quantum regressor to pre-trained MLP or GNN embeddings and compare three regimes: training from scratch (HQSc), fine-tuning the entire pre-trained hybrid (HQFi), and freezing the classical feature extractor while training only the quantum layer (HQFr) (Cho et al., 1 Mar 2025).

Several papers depart from pure backpropagation to address hybrid-specific bottlenecks. GAT-QNN introduces a two-stage genetic algorithm: a macroCircuit is trained by sampling and training microCircuits whose learned parameters are reintegrated into the macroCircuit, then a separate GA-driven inference stage selects backend-specific microCircuits without retraining (Ahmed et al., 16 Apr 2026). The qtDNN method for reinforcement learning uses a local classical surrogate that approximates the PQC from buffered input–output pairs; the surrogate replaces the PQC inside the computation graph during batched backpropagation, while the real quantum layer is retained for inference (Luo et al., 12 Mar 2025). A more radical formulation casts quantized neural-network training as QCBO and solves it with Quantum Conditional Gradient Descent, avoiding large penalty-coefficient tuning and using a coherent Ising machine as the QUBO oracle (Li et al., 23 Jun 2025).

These alternatives indicate that “training a QT-NN” is not synonymous with differentiating through a fixed ansatz. The literature includes shared-weight evolution, local surrogate replacement, and quantum-oracle optimization, all under the same hybrid objective of coupling quantum structure to classical model learning (Ahmed et al., 16 Apr 2026, Luo et al., 12 Mar 2025, Li et al., 23 Jun 2025).

4. Compression, inference locality, and resource allocation

A defining thread in QT-NN research is the relocation of quantum cost from inference to training. Quantum-Train states this explicitly: it reduces the number of trainable parameters from $h_{\phi}$1 to $h_{\phi}$2 during training, avoids quantum data encoding of input samples, and leaves a fully classical neural network for deployment (Liu et al., 2024). This is not merely an engineering convenience. It changes the role of the quantum module from predictor to parameter prior, adapter, or hypernetwork.

QT-CNN gives a concrete instance of this regime. Its QNN blocks use layered $h_{\phi}$3 rotations and CNOT chains on $h_{\phi}$4 qubits to represent $h_{\phi}$5 classical parameters, while the mapping model and scaling model remain compact. On the DEEP-VOICE task, the classical baseline has 3,373 parameters, whereas QT-CNN spans 456–1,464 total parameters depending on the number of QNN blocks, amounting to 13.5–43.4% of the classical model; the framework reports up to 70% fewer trainable parameters while matching or closely tracking the classical baseline in training and test accuracy (Lin et al., 2024).

Tensor-network and photonic variants push this compression logic further. Replacing the original MLP mapper with an MPS improves scalability and interpretability, and the distributed circuit ansatz factorizes a large quantum state into smaller subcircuits whose measurement statistics are recombined classically; on CIFAR-10, a QT-TN system reached approximately 60% test accuracy with slightly over 10,000 parameters, whereas the MLP-based mapper required approximately 23,000 parameters at comparable accuracy for a target CNN with 285,226 parameters (Liu et al., 2024). In distributed photonic QT, photonic QNNs with $h_{\phi}$6 trainable interferometer parameters generate photon-count distributions that are mapped to classical weights via an MPS. On MNIST, this achieved $h_{\phi}$7 accuracy with 3,292 parameters at $h_{\phi}$8, compared with $h_{\phi}$9 for a classical baseline with 6,690 parameters; at $f_{\theta}$0, a ten-fold compression ratio was reported with a relative accuracy loss of less than $f_{\theta}$1, and replacing the photonic QNN with random inputs reduced performance to chance level, $f_{\theta}$2 (Chen et al., 13 May 2025).

Resource allocation is also treated as a tunable classical–quantum boundary. The MPS–VQC classifier is explicitly described as adaptable: the feature dimension output by the tensor network can be matched to available qubits, and part or all of the MPS may be re-implemented as a quantum circuit as quantum resources grow (Chen et al., 2021). GAT-QNN adds a deployment-time variant of the same idea: smaller microCircuits can be selected after training to reduce gate count and improve backend-specific performance without retraining (Ahmed et al., 16 Apr 2026).

5. Empirical evidence across tasks

On image classification, end-to-end hybridization has repeatedly outperformed simple classical compression baselines. For MNIST $f_{\theta}$3 vs $f_{\theta}$4, PCA-VQC reduced 784-dimensional inputs to four principal components and reached $f_{\theta}$5 testing accuracy with testing loss $f_{\theta}$6, whereas hybrid MPS-VQC with bond dimension $f_{\theta}$7 reached $f_{\theta}$8 testing accuracy with testing loss $f_{\theta}$9; the standalone MPS with $g_{\phi}$0 overfit, but the hybrid did not show rising test loss, leading the authors to describe the VQC as a regularizer (Chen et al., 2020). On broader benchmark classification, Quantum-Train traded some accuracy for compression: the classical CNN reached $g_{\phi}$1 on MNIST, $g_{\phi}$2 on FashionMNIST, and $g_{\phi}$3 on CIFAR-10, whereas QT-NN configurations reported $g_{\phi}$4, $g_{\phi}$5, and $g_{\phi}$6, respectively, with much smaller parameter counts (Liu et al., 2024).

Hybrid QT-NNs have also been evaluated in audio and security settings. QT-CNN for deepfake-audio detection reported comparable performance to traditional CNNs while reducing trainable parameters by up to 70% (Lin et al., 2024). In UAV swarm intrusion detection, QT-NNs were benchmarked against quantum kernels, QNNs, and strong classical baselines under matched preprocessing and noise assumptions; the reported conclusion was that quantum kernels and QT-NNs excel in low-data, nonlinear regimes, deeper QNNs suffer trainability issues, and CNNs dominate when abundant data offset their larger parameter count (Chen et al., 1 Sep 2025).

For regression and scientific modeling, the record is mixed but nontrivial. On the Boston Housing dataset, a sandwiched hybrid QNN slightly outperformed a classical NN with identical settings and showed lower test loss, but training time was about three times longer, and the photonic continuous-variable model failed to generalize under severe simulation constraints (Jain et al., 2022). In QSPR modeling of $g_{\phi}$7-capturing amines, HQNNs improved predictive accuracy for basicity, viscosity, boiling point, melting point, and vapor pressure; the fine-tuned and frozen pre-trained 9-qubit models were ranked highest, and simulated IBM noise left predictive performance intact, with slightly improved $g_{\phi}$8 and MAE reported for basicity in one noisy setting (Cho et al., 1 Mar 2025). In seismic inversion, HQ-PINN consistently estimated accurate P- and S-impedances on synthetic and Sleipner field data, and even the “no ansatz” quantum node performed nearly as well as trainable entangler variants in post-stack inversion (Vashisth et al., 6 Mar 2025).

Quantum-enhanced training has also been explored in domains where the task is itself quantum or sequentially decision-based. For quantum many-body simulation, a hybrid ansatz combining autoregressive neural quantum states with PQCs achieved lower relative energies than standalone NQS; on LiH in a 6-qubit STO-3G basis, the hybrid reported an average error of $g_{\phi}$9 Ha, compared with $N$0 Ha for a same-sized NQS and $N$1 Ha for a 10x-larger NQS (Zhang et al., 21 Jan 2025). In reinforcement learning, hDQNN-TD3 on Humanoid-v4 surpassed classical TD3, SAC, and PPO baselines trained with identical compute, and was presented as the first PQC-enhanced policy to match or exceed state-of-the-art classical performance on that benchmark (Luo et al., 12 Mar 2025).

6. Trainability, limitations, and current points of contention

A recurrent assumption in PQC analysis is an expressibility–trainability trade-off: more expressive circuits should be harder to optimize because gradients vanish. Recent work argues that this heuristic does not transfer cleanly to hybrid models. In a systematic study of circuit depth, qubit count, and entanglement topology, pure PQCs exhibited only a weak, regime-dependent trade-off, while fully trainable hybrid models eliminated the relation under end-to-end optimization; classical preprocessing and postprocessing were found to reshape the effective optimization landscape, motivating multi-objective neural architecture search over the joint classical–quantum space rather than quantum-only heuristics (Kashif et al., 25 May 2026). This directly challenges the misconception that hybridization is a superficial wrapper around PQCs.

A second misconception is that QT-NNs necessarily require quantum hardware at inference. QT, QT-CNN, tensor-network-mapped QT, and photonic-QT all explicitly decouple training from deployment, leaving a classical model after training (Liu et al., 2024, Lin et al., 2024, Chen et al., 13 May 2025). A third is that stronger quantum parameterization automatically yields better models. TunnElQNN reported improvement as quantum-layer depth increased up to four layers, followed by a slight performance drop attributed to overfitting or vanishing gradients (Abbas, 2 May 2025), while recent UAV intrusion-detection results similarly associated deeper QNNs with trainability problems (Chen et al., 1 Sep 2025).

Hardware constraints remain decisive. The neuron-as-VQC binary classifier achieved roughly 10% higher classification accuracy and 20% better cost minimization than an individual VQC on simulated hardware, yet on real IBM devices both models performed well only when qubit and gate counts were sufficiently small (Arthur et al., 2022). Continuous-variable photonic QNNs were described as computationally infeasible beyond trivial feature counts in classical simulation (Jain et al., 2022). Seismic inversion experiments were conducted on noiseless simulators, with real-hardware effects left for future work (Vashisth et al., 6 Mar 2025). Even when noise robustness is reported—as in QSPR on IBM-inspired noise models or photonic QT under realistic photonic imperfections—the results remain task- and architecture-dependent rather than universal (Cho et al., 1 Mar 2025, Chen et al., 13 May 2025).

The present literature therefore supports a restrained interpretation. Hybrid QT-NNs have demonstrated at least four concrete capabilities: end-to-end co-optimization of classical and quantum modules, parameter-efficient weight generation, deployment-time classical inference in QT-style systems, and application-specific gains in classification, regression, scientific modeling, and reinforcement learning. At the same time, their benefits are contingent on architecture choice, training regime, problem structure, and hardware limits. This suggests that QT-NNs are best understood not as a single model class with settled advantages, but as a design space in which the quantum component may serve as feature map, classifier, regularizer, hypernetwork, optimizer, or distributed resource generator, with the classical component often determining whether those roles become trainable and practically useful (Kashif et al., 25 May 2026).