Quantum Adaptive Excitation Network (QAE-Net)

• QAE-Net is a quantum-enhanced attention module that replaces the classical excitation MLP with a shallow variational quantum circuit while retaining standard squeeze and reweight functions.
• The framework encodes channel descriptors into quantum states, measures Pauli‐Z expectation values, and employs the parameter‐shift rule for seamless end-to-end CNN training.
• Empirical evaluations show notable accuracy gains on datasets like CIFAR-10, demonstrating the effectiveness of integrating quantum components into classical CNN architectures.

Searching arXiv for the QAE-Net paper and closely related quantum autoencoder / adaptive excitation work. Quantum Adaptive Excitation Network (QAE-Net) is a hybrid quantum-classical framework for channel attention in Convolutional Neural Networks (CNNs). In its published form, QAE-Net preserves the squeeze-excitation-reweight pattern of Squeeze-and-Excitation networks, but replaces the classical excitation multilayer perceptron with a shallow Variational Quantum Circuit (VQC). The module performs global average pooling to form a channel descriptor, encodes that descriptor into a quantum state, measures Pauli-$Z$ expectation values, and maps the quantum outputs back to channel-wise attention coefficients for feature recalibration. Empirically, the architecture was evaluated on MNIST, FashionMNIST, and CIFAR-10, with consistent gains over a matched SENet baseline and especially large improvements on the three-channel CIFAR-10 setting; increasing the number of variational layers from 1 to 3 further raised CIFAR-10 accuracy from $89.08\%$ to $92.30\%$ (Hsu et al., 15 Jul 2025).

1. Definition, scope, and nomenclature

QAE-Net is defined as a channel-attention module rather than as a quantum autoencoder. Its role is to sit inside a CNN at the same architectural location as an SE block and to replace the excitation stage with a VQC, while keeping the squeeze operation and channel-wise reweighting classical. The explicit design target is “channel attention mechanisms in Convolutional Neural Networks,” and the stated motivation is to use quantum superposition and entanglement to capture higher-order inter-channel dependencies that are challenging for small classical excitation MLPs to model (Hsu et al., 15 Jul 2025).

The acronym is adjacent to, but distinct from, the “QAE” literature on quantum autoencoders. In that literature, a QAE typically compresses information by partitioning qubits into latent and trash registers, tracing out the trash subsystem, and reconstructing the input through an inverse or decoder circuit. Representative examples include a fully quantum autoencoder for 3D point clouds (Rathi et al., 2023) and QCNN-based quantum autoencoders for image feature extraction and reconstruction (Lo et al., 11 Feb 2025). QAE-Net does not use that compression–reconstruction formalism; its published instantiation is an attention subroutine embedded in a classical vision backbone.

This distinction matters because two separate research trajectories are involved. One trajectory studies quantum autoencoding as compression and reconstruction. The other, exemplified by QAE-Net, studies a narrowly scoped quantum replacement for a classical architectural primitive. A plausible implication is that QAE-Net is best understood as a modular quantum attention mechanism rather than as a general-purpose latent-variable model.

2. Architectural formulation

Given a feature map $\mathbf{X} \in \mathbb{R}^{C \times H \times W}$, QAE-Net first performs the standard squeeze operation through global average pooling: $$\mathbf{z}_c = \frac{1}{H \times W} \sum_{i=1}^{H} \sum_{j=1}^{W} \mathbf{X}_{c,i,j}, \quad c = 1,\dots,C.$$ This produces the channel descriptor $\mathbf{z} \in \mathbb{R}^C$, exactly as in SENet. The novelty begins at the excitation stage, where $\mathbf{z}$ is grouped into triples of scalars, each triple feeding one qubit. The register is initialized in $|0\rangle^{\otimes n}$, acted on by Hadamards, and then angle-encoded through

$$U_{\text{enc}}(\mathbf{x}) = \bigotimes_{i=1}^{n} RZ(x_{3i-2}) \cdot RY(x_{3i-1}) \cdot RZ(x_{3i}).$$

The encoded state is

$$|\phi(\mathbf{z})\rangle = U_{\text{enc}}(\mathbf{z}) \, H^{\otimes n} |0\rangle^{\otimes n}.$$

In the reported experiments, $89.08\%$0 qubits were used throughout (Hsu et al., 15 Jul 2025).

The variational stage applies $89.08\%$1 layers of parameterized single-qubit rotations and entangling CNOTs: $89.08\%$2 with the entangling pattern written as

$89.08\%$3

The full quantum state is therefore

$89.08\%$4

After the VQC, QAE-Net measures the expectation of $89.08\%$5 on each qubit,

$89.08\%$6

forming the quantum feature vector $89.08\%$7. A classical linear map with sigmoid then returns to channel dimension: $89.08\%$8 and the attention coefficients rescale the original feature map channel-wise: $89.08\%$9 The module is therefore structurally conservative at the CNN level and quantum only at the excitation stage (Hsu et al., 15 Jul 2025).

In the reported backbone, the surrounding network was

$92.30\%$0

The only architectural difference between SENet and QAE-Net was the excitation submodule.

3. Quantum excitation, optimization, and trainability

The end-to-end objective is standard multiclass cross-entropy. If the downstream classifier produces logits $92.30\%$1, the softmax probabilities are

$92.30\%$2

and the minibatch loss is

$92.30\%$3

Gradients through the quantum layer are computed by the parameter-shift rule. For an observable $92.30\%$4,

$92.30\%$5

This lets the VQC participate in ordinary joint training with the classical CNN components (Hsu et al., 15 Jul 2025).

The implementation uses PennyLane for the quantum circuit and automatic differentiation, PyTorch for the classical network, and the lightning.gpu backend for statevector simulation. The hardware platform reported for simulation was NVIDIA Tesla V100 GPUs with Intel Xeon CPUs. All experiments were performed in simulation rather than on quantum hardware (Hsu et al., 15 Jul 2025).

Trainability is treated as a central practical issue. The discussion explicitly notes the risk of barren plateaus in deeper or wider VQCs and presents QAE-Net’s mitigation strategy as architectural containment: four qubits, one to three variational layers, and use of the quantum circuit only inside a small attention routine. This is consistent with broader hybrid QML practice, where shallow parameter-shift-compatible PQCs are embedded in end-to-end optimization loops rather than used as full-network replacements (Zhang et al., 2024).

The “excitation” terminology is literal at the architectural level. In classical SE networks, the excitation block is a low-dimensional gating MLP. In QAE-Net, excitation is realized by data-dependent rotations, variational single-qubit rotations, and entangling CNOTs whose measured expectation values are then remapped to attention gates. The resulting nonlinearity is therefore hybrid: unitary evolution plus measurement-induced classical interface.

4. Empirical performance and scaling with variational depth

The central reported comparison is against SENet with the same backbone and nearly identical parameter count. The published results are as follows (Hsu et al., 15 Jul 2025).

Setting	Params	Test Acc. (%)
MNIST SENet	39,602	97.9
MNIST QAE-Net	39,366	98.0
FashionMNIST SENet	39,602	91.0
FashionMNIST QAE-Net	39,366	91.3
CIFAR-10 SENet	142,634	76.72
CIFAR-10 QAE-Net	142,570	89.08

Two empirical regimes are visible. On MNIST and FashionMNIST, both single-channel datasets, QAE-Net improves test accuracy modestly while slightly reducing parameter count. On CIFAR-10, which has three-channel RGB inputs, the improvement is much larger: $92.30\%$6 for SENet versus $92.30\%$7 for QAE-Net, with a parameter difference of only 64. This aligns with the paper’s emphasis that gains are particularly notable on tasks involving three-channel inputs (Hsu et al., 15 Jul 2025).

A second key result concerns depth. On CIFAR-10, with the same four qubits and learning rate $92.30\%$8, increasing the number of variational layers monotonically improved accuracy:

• QAE-Net (1 layer): $92.30\%$9
• QAE-Net (2 layers): $\mathbf{X} \in \mathbb{R}^{C \times H \times W}$0
• QAE-Net (3 layers): $\mathbf{X} \in \mathbb{R}^{C \times H \times W}$1

The training curves are described as showing that QAE-Net converges faster and reaches lower final cross-entropy loss than SENet on all three datasets. The depth study likewise reports progressively better loss reduction as the number of variational layers increases. In the paper’s interpretation, this supports the claim that deeper VQCs are more expressive and can capture more complex inter-channel dependencies (Hsu et al., 15 Jul 2025).

The most cautious reading is architectural rather than ontological. The results establish that, in this small-scale simulated setting, replacing the excitation MLP with a 4-qubit VQC can improve a matched CNN. They do not by themselves establish a general quantum advantage across attention mechanisms, datasets, or hardware conditions.

5. Relation to quantum autoencoders and adjacent hybrid architectures

Although QAE-Net is not a quantum autoencoder, it sits near several active lines of quantum neural-network design. One is the Romero-style quantum autoencoder tradition, in which qubits are partitioned into compressed and trash registers, compression is implemented by partial trace, and reconstruction is performed by an inverse or decoder circuit. In a QCNN-based study of unsupervised feature extraction and reconstruction, an 8-qubit encoder compressed to 1 qubit and achieved an average accuracy of $\mathbf{X} \in \mathbb{R}^{C \times H \times W}$2 for binary MNIST classification of digits 0 and 1; that work also reported that MSE outperformed BCE and that QCNN structure outperformed more universal ansätze (Lo et al., 11 Feb 2025). Another example, “3D-QAE,” introduced a fully quantum autoencoder for 3D point clouds, including amplitude encoding, a bottleneck induced by tracing out qubits, and a decoder circuit on simulated gate-based quantum hardware (Rathi et al., 2023).

A second adjacent line is the hybrid autoencoder tradition in which only part of the encoder is quantum. In an end-to-end communication system, a hybrid quantum-classical autoencoder replaced the transmitter with two parallel PQCs and kept a classical receiver; it reported BLER comparable to classical and conventional schemes while significantly reducing trainable parameters (Zhang et al., 2024). A third line uses neural architecture search for hybrid quantum-classical autoencoders, where a genetic algorithm evolves VQC gate patterns and entanglement structure to reduce reconstruction loss under fixed qubit and depth budgets (Agha et al., 24 Nov 2025).

These lines do not define QAE-Net, but they clarify its architectural neighborhood. All of them rely on the same interface pattern: a low-dimensional classical summary is encoded into a shallow quantum circuit, measurement returns a classical vector, and the surrounding learning system remains classical. This suggests that QAE-Net belongs to a broader family of quantum subroutine architectures in which the quantum component is not the full model but a narrowly targeted operator for compression, attention, or encoding.

6. Limitations, open problems, and research directions

The published limitations of QAE-Net are explicit. All results were obtained in simulation; real NISQ devices introduce noise, limited fidelity, connectivity constraints, and shot noise. The reported experiments used small CNNs, only four qubits, and only one to three variational layers. Scaling to large channel counts and larger images is unresolved, and the current encoding—three real numbers per qubit via $\mathbf{X} \in \mathbb{R}^{C \times H \times W}$3-$\mathbf{X} \in \mathbb{R}^{C \times H \times W}$4-$\mathbf{X} \in \mathbb{R}^{C \times H \times W}$5—is identified as potentially suboptimal when $\mathbf{X} \in \mathbb{R}^{C \times H \times W}$6 becomes large. The paper also notes unresolved issues of latency and throughput for real-hardware deployment (Hsu et al., 15 Jul 2025).

The paper’s proposed directions are concrete: problem-specific circuit architectures, more qubits and deeper circuits with mitigation of barren plateaus, better encoding schemes, extension from channel attention to spatial attention, and transfer to other vision tasks such as segmentation and detection, as well as non-vision domains such as time-series data. It also recommends layerwise or blockwise training and local cost functions as future responses to trainability limits (Hsu et al., 15 Jul 2025).

Related work tempers the depth narrative. QAE-Net reports monotonic gains from one to three variational layers on four-qubit CIFAR-10. However, QCNN-based quantum autoencoder experiments on MNIST reported that a 16-qubit QCNN with four convolution-pooling layers performed worse than an 8-qubit QCNN with three such layers (Lo et al., 11 Feb 2025), and genetic-search studies of hybrid autoencoders found that additional entanglement often improved reconstruction but that excessive entanglement could increase loss by roughly $\mathbf{X} \in \mathbb{R}^{C \times H \times W}$7 to $\mathbf{X} \in \mathbb{R}^{C \times H \times W}$8 (Agha et al., 24 Nov 2025). This suggests that the QAE-Net depth ablation should be read as a local result for a specific small circuit class, not as a universal prescription that deeper quantum attention blocks are always preferable.

A persistent misconception is that replacing a small classical MLP with a VQC automatically yields a compression or expressivity gain that dominates classical alternatives. The current evidence is narrower. QAE-Net demonstrates that a shallow 4-qubit excitation module can outperform a matched SE block on the reported datasets, especially on CIFAR-10. It does not eliminate the standard constraints of hybrid QML: simulator dependence, measurement overhead, architecture sensitivity, and trainability deterioration as qubit count, entanglement density, or depth increase.

From the current literature, the most durable interpretation is modular. QAE-Net is a quantum-enhanced attention block, not a universal replacement for CNNs, not a synonym for quantum autoencoders, and not yet a hardware-validated primitive. Its significance lies in showing that a small variational circuit can be inserted into a standard deep-learning workflow at the excitation stage of channel attention and can alter the accuracy–capacity trade-off in a measurable way under NISQ-compatible circuit budgets (Hsu et al., 15 Jul 2025).