Hybrid Quantum-Classical Neural Networks

• Hybrid quantum-classical neural networks are architectures that integrate parameterized quantum circuits with classical layers to exploit quantum feature maps and enhance learning.
• They employ innovative techniques such as angle and amplitude encoding, variational quantum circuits, and joint optimization to achieve faster convergence and improved accuracy.
• Empirical studies reveal benefits including reduced epochs to convergence and 1–1.5% accuracy gains in applications like image classification and time-series forecasting.

Hybrid quantum-classical neural network architectures integrate quantum circuits with classical neural network modules to exploit the representational power of quantum feature maps and classical computation. These architectures, designed for implementation on near-term Noisy Intermediate-Scale Quantum (NISQ) hardware, span various domains, from image classification to time-series forecasting and control. Hybrid models commonly demonstrate improved convergence properties and, in specific contexts, outperform their purely classical analogues. Their construction involves precise circuit design, principled data encoding, a co-optimization of classical and quantum parameters, and explicit consideration of quantum resource limitations.

1. Architectural Principles and Model Variants

Hybrid quantum-classical neural networks comprise a compositional pipeline where quantum and classical layers are arranged either sequentially, in alternating patterns, or in parallel. Typical models include:

1. Quantum-Classical Convolutional Neural Networks (QCCNN, QCResNet): Classical input data is processed through quantum convolutional layers, with each filter realized as a parameterized quantum circuit acting on compact local patches (e.g., $2 \times 2$ or $3 \times 3$), followed by classical layers for downstream feature extraction and classification. Hybrid residual architectures replace designated classical convolutions within standard ResNet blocks with quantum circuits (Shi et al., 2023).
2. Non-Sequential/Alternating Layer Models (TunnElQNN): Alternation of classical and quantum blocks, with the classical layers using physics-inspired nonlinearities (such as the Tunneling-Diode Activation Function, TDAF) to modulate both pre- and post-processing relative to quantum operations. These models can present skip connections and are evaluated for expressiveness under varying quantum depth (Abbas, 2 May 2025).
3. Classical-to-Quantum Transfer Learning: A pre-trained deep classical network (e.g., ResNet-18) is truncated near the classification head, and its features are embedded into a low-dimensional quantum circuit for final classification. The classical weights are typically frozen, with only the quantum module and shallow classical layers jointly trained (Mari et al., 2019).
4. Parallel Hybrid Networks: Input data is simultaneously processed by both a variational quantum circuit and a classical MLP, before the two outputs are fused by a trainable linear combination. This architecture leverages quantum modules for representing smooth periodic components and classical nets for non-harmonic corrections (Kordzanganeh et al., 2023).
5. Hybrid Quantum Feature Extraction: Quantum modules (e.g., QuFeX) act as bottleneck feature extractors, inserted at chosen locations in architectures such as U-Net, providing a reduced-dimensional transformation that encodes non-local correlations (Jain et al., 22 Jan 2025).
6. Hybrid Recurrent Networks: Parameterized quantum circuits (PQC) act as the recurrent core, driven by a classical controller network that supplies gate parameters based on input and mid-circuit quantum measurements. Mid-circuit readouts provide nonlinearities and enable attention mechanisms (Xu, 29 Oct 2025).

2. Quantum Layer Construction and Data Encoding

Quantum layers in hybrid architectures are typically based on variational quantum circuits parameterized by a set of continuous rotation angles, entangling gates, and measurement observables. Key design elements:

• Data encoding (Feature Maps):
  • Angle encoding: Each classical feature $x_i$ is mapped to a qubit via a rotation, typically $R_y(x_i)$ or a product of $R_x$, $R_y$, $R_z$ (e.g., AngleEmbedding with $U(x_i)=R_X(x_i)R_Y(x_i)R_Z(x_i)$) (Shi et al., 2023, Abbas, 2 May 2025).
  • Amplitude encoding: Classical vectors are normalized and loaded as amplitudes into the quantum state $|\psi\rangle=\sum x_i|i\rangle$ (Anwar et al., 25 Aug 2025).
  • Hybrid transfer: Classical pre-processing maps high-dimensional features down to $n_q$ via a dense layer and $tanh$, then encodes as $R_y(\pi x_i)$ (Mari et al., 2019).
• Parameterized variational block:
  • Single/multi-layer ansatz: Blocks alternate single-qubit rotations (e.g., $R_x$, $R_y$, $R_z$) with all-to-all or nearest-neighbour entangling gates (e.g., CNOT ladder or ring) (Shi et al., 2023).
  • Entanglement pattern: The expressivity is controlled by circuit depth and entanglement; strong vs. basic entangling layouts directly affect model accuracy (Zaman et al., 2024).
• Measurement and classical mapping: Qubit expectation values (e.g., $\langle Z_i\rangle$) are used as features in downstream classical layers and appropriately post-processed (e.g., $(z_i + 1)/2$ mapped to $[0,1]$ per channel).

3. Optimization and Training Procedures

Training of hybrid networks involves joint optimization of quantum circuit parameters (e.g., rotation angles, entangling gate strengths) and classical weights (linear/dense layers, biases). This process is enabled by:

• Loss functions:
  • Multi-class cross-entropy is standard for classification tasks (Shi et al., 2023).
  • Regression networks (e.g., financial forecasting) use mean squared error (Choudhary et al., 19 Mar 2025).
  • Fidelity-based loss (SWAP test) for overlap between data-encoded and class states (Stein et al., 2021).
• Gradient estimation:
  • Parameter-shift rule: For any quantum parameter $\theta_j$, the gradient is evaluated as

  $$\frac{\partial L}{\partial\theta_j} = \frac{L(\theta_j+\pi/2) - L(\theta_j-\pi/2)}{2}$$

  for observables with eigenvalues $\pm1$ (Shi et al., 2023, Abbas, 2 May 2025, Dehaghani et al., 2024). - Quantum gradients are propagated through the quantum-classical computational graph via autodiff frameworks where possible.

• Optimization algorithms: Adam (Abbas, 2 May 2025, Jain et al., 22 Jan 2025, Austin et al., 2024) and SGD (Shi et al., 2023), with learning rates empirically tuned ($10^{-3}$–$10^{-2}$).

• Shot sampling: Quantum expectations are estimated from $S$ circuit evaluations (e.g., $S=1500$ found optimal (Shi et al., 2023)).

4. Performance, Scalability, and Empirical Evaluation

Empirical results on hybrid quantum-classical neural networks indicate:

• Faster convergence: Hybrid models typically converge much faster than their classical counterparts, e.g., 30–40 epochs vs. $>$100 for classical CNN (Shi et al., 2023).

• Accuracy gains: QCCNN/QCResNet outperform analogous classical architectures by $\sim$1--1.5% accuracy on image tasks (Shi et al., 2023).

• Robustness to class overlap: Hybrid architectures with non-standard activations maintain test accuracy $>90\%$ under high class overlap, where classical models degrade substantially (Abbas, 2 May 2025).

• Scalability considerations:

  • Qubit count: Kept $<$10 (usually $4$ or $9$) by operating on small patches or strongly compressing features (Shi et al., 2023, Mari et al., 2019).
  • Circuit depth: Limited to 1–2 variational layers to match NISQ coherence constraints and suppress barren plateaus (Shi et al., 2023, Alam et al., 2022).
  • Feature compression: Classical layers reduce the dimensionality presented to the quantum device, enabling hybridization even for high-dimensional data (Mari et al., 2019).
• Empirical summary (extracted examples):

| Model | Dataset | Test Acc (%) | Epochs to Convergence | |---------------|------------------|------------------|----------------------| | QCCNN-1 | Phytoplankton | 93 | 30–40 | | QCResNet-1 | Phytoplankton | 93.2 | 30–40 | | CNN | Phytoplankton | 92 | >100 | | ResNet | Phytoplankton | 91.9 | >100 | | TunnElQNN | Synthetic (Δ=1.5)| 97 | 150 | | ReLUQNN | Synthetic (Δ=1.5)| 87 | 150 |

Hybrid networks also exhibit resilience to quantum noise if constructed with shallow, modular quantum layers (Alam et al., 2022). For image classification and segmentation, quantum feature extractors or bottleneck modules (QuFeX) placed at information bottlenecks have shown to improve segmentation accuracy and reduce parameter count compared to classical baselines once total size exceeds $\sim$25k parameters (Jain et al., 22 Jan 2025).

5. Design Trade-Offs and Model Selection

Designing effective hybrid quantum-classical neural networks requires careful trade-off balancing:

• Shallow quantum circuits: Shallow depth counters noise accumulation and avoids barren plateaus; deeper quantum circuits provide increased expressiveness but are sensitive to hardware limitations and gradient vanishing (Alam et al., 2022, Shi et al., 2023, Abbas, 2 May 2025).
• Hybridization placement: Early-layer quantum convolutions can replace classical feature extractors, while deep classical layers manage large-scale spatial hierarchies (Shi et al., 2023). Bottleneck insertion (e.g., U-Net) and parallel fusion are alternative strategies (Jain et al., 22 Jan 2025, Kordzanganeh et al., 2023).
• Quantum ansatz expressiveness: Expressibility and entanglement must be tuned to provide learning capacity without rendering the landscape untrainable (Shi et al., 2023, Zaman et al., 2024).
• Encoding choice: Amplitude encoding is more powerful but more resource-intensive, while angle encoding is more NISQ-feasible (Freinberger et al., 8 Jan 2026). Data normalization and scaling are crucial for optimal circuit input.

Guidelines for practitioners, as drawn from empirical studies:

• Use compact quantum filters (patches of $2\times2$ or $3\times3$) and few qubits per circuit (Shi et al., 2023).
• Hybridize only early feature-extraction or bottleneck layers to avoid qubit proliferation (Jain et al., 22 Jan 2025).
• Employ TDAF or other physics-inspired activations to enrich classical-quantum nonlinearities (Abbas, 2 May 2025).
• Systematically ablate circuit depth, entanglement, and classical-quantum partitioning to tune performance (Zaman et al., 2024).

6. Application Domains and State-of-the-Art Tasks

Hybrid quantum-classical neural networks have been demonstrated in:

• Image classification: QCCNN, QCResNet, TunnElQNN, QuanNN, QuFeX-U-Net achieve state-of-the-art or near-optimal performance on MNIST, Fashion-MNIST, OrganAMNIST, and phytoplankton datasets (Shi et al., 2023, Abbas, 2 May 2025, Anwar et al., 25 Aug 2025, Jain et al., 22 Jan 2025).
• Time-series forecasting: Hybrid architectures with classical RNNs/LSTMs for temporal modeling and quantum circuits for nonlinear mapping yield performance advantages in stock market regression under both sequential and joint optimization (Choudhary et al., 19 Mar 2025).
• Recurrent and sequence learning: Quantum recurrent cores with explicit classical control (QRNN) achieve competitive performance to LSTM and scoRNN on sentiment analysis, sequence modeling, and translation (Xu, 29 Oct 2025).
• Physics-informed learning and optimal control: Quantum-classical PINN frameworks, where quantum neural networks approximate free functions in the variational ansatz, accelerate convergence and state-to-target fidelity in quantum control landscapes (Dehaghani et al., 2024).

7. Assessment and Outlook

While empirical and benchmark studies demonstrate that hybrid quantum-classical neural network architectures can achieve faster convergence and, in some configurations, superior accuracy and robustness relative to equivalent classical models, recent systematic evaluations highlight critical caveats:

• Quantum components contribute positively to performance only in select scenarios and may degrade accuracy—especially on high-dimensional or 3D inputs—if not carefully designed and placed (Freinberger et al., 8 Jan 2026).
• Amplitude encoding generally yields higher representational capacity but may be infeasible due to quantum data-loading costs in real hardware (Freinberger et al., 8 Jan 2026).
• Entanglement and circuit size should be scaled commensurately with classical feature compression; overparameterization or insufficiently compressed latent vectors can quickly erode hybrid performance.
• Robust benchmarking using nonparametric statistical tests against matched classical baselines is essential to substantiate any claims of quantum advantage in current NISQ applications (Freinberger et al., 8 Jan 2026).

Anticipated future advances include enhanced hardware-efficient circuit design, error-mitigation tailored to modular hybrid architectures, and extension to domains such as quantum optimal control, segmentation, and generative modeling.

---

References

• "Hybrid quantum-classical convolutional neural network for phytoplankton classification" (Shi et al., 2023)
• "TunnElQNN: A Hybrid Quantum-classical Neural Network for Efficient Learning" (Abbas, 2 May 2025)
• "Transfer learning in hybrid classical-quantum neural networks" (Mari et al., 2019)
• "Hybrid Quantum-Classical Learning for Multiclass Image Classification" (Anwar et al., 25 Aug 2025)
• "A Hybrid Quantum-Classical Neural Network Architecture for Binary Classification" (Arthur et al., 2022)
• "DeepQMLP: A Scalable Quantum-Classical Hybrid DeepNeural Network Architecture for Classification" (Alam et al., 2022)
• "HQNN-FSP: A Hybrid Classical-Quantum Neural Network for Regression-Based Financial Stock Market Prediction" (Choudhary et al., 19 Mar 2025)
• "Hybrid Quantum-Classical Recurrent Neural Networks" (Xu, 29 Oct 2025)
• "A Hybrid Quantum-Classical Physics-Informed Neural Network Architecture for Solving Quantum Optimal Control Problems" (Dehaghani et al., 2024)
• "An end-to-end trainable hybrid classical-quantum classifier" (Chen et al., 2021)
• "QuClassi: A Hybrid Deep Neural Network Architecture based on Quantum State Fidelity" (Stein et al., 2021)
• "Quantum feature extraction module for hybrid quantum-classical deep neural networks" (Jain et al., 22 Jan 2025)
• "Parallel Hybrid Networks: an interplay between quantum and classical neural networks" (Kordzanganeh et al., 2023)
• "Lean classical-quantum hybrid neural network model for image classification" (Liu et al., 2024)
• "Hybrid Quantum-Classical Photonic Neural Networks" (Austin et al., 2024)
• "A Comparative Analysis of Hybrid-Quantum Classical Neural Networks" (Zaman et al., 2024)
• "Enhanced image classification via hybridizing quantum dynamics with classical neural networks" (Zhou et al., 18 Jul 2025)
• "Leveraging Quantum Layers in Classical Neural Networks" (Illésová, 16 Jul 2025)
• "The Role of Quantum in Hybrid Quantum-Classical Neural Networks: A Realistic Assessment" (Freinberger et al., 8 Jan 2026)