Layered Quantum Architecture Search for 3D Point Cloud Classification

Natacha Kuete Meli; Jovita Lukasik; Vladislav Golyanik; Michael Moeller

quant-phMar 20, 2026

Layered Quantum Architecture Search for 3D Point Cloud Classification

Authors: Natacha Kuete Meli, Jovita Lukasik, Vladislav Golyanik, Michael Moeller

Organizations: University of Siegen · MPI for Informatics, SIC

Abstract

We introduce layered Quantum Architecture Search (layered-QAS), a strategy inspired by classical network morphism that designs Parametrised Quantum Circuit (PQC) architectures by progressively growing and adapting them. PQCs offer strong expressiveness with relatively few parameters, yet they lack standard architectural layers (e.g., convolution, attention) that encode inductive biases for a given learning task. To assess the effectiveness of our method, we focus on 3D point cloud classification as a challenging yet highly structured problem. Whereas prior work on this task has used PQCs only as feature extractors for classical classifiers, our approach uses the PQC as the main building block of the classification model. Simulations show that our layered-QAS mitigates barren plateau, outperforms quantum-adapted local and evolutionary QAS baselines, and achieves state-of-the-art results among PQC-based methods on the ModelNet dataset.

Figures & tables

Figure 1: Overview of our framework for 3D point classification using Parametrised Quantum Circuits (PQCs). The input point cloud is voxelised and used to prepare the quantum system. We then use the new layered Quantum Architecture Search (layered-QAS) approach to engineer the PQC design that meaningfully learns features from the encoded point cloud. Lastly, qubits are measured to extract the learned features, which are used by an optionally learnable classical linear layer for classification.

Figure 2: Amplitude encoding of a 3D point cloud into a quantum state. We partition the point cloud into voxels (left) and compute the normalised density of each voxel, i.e . the proportion of points within the voxel (middle). This creates a matrix that we vectorise and use as input state vector for the PQC (right). The granularity of the voxelisation, which defines the number of qubits used, is user-defined. Colours are added for visualisation purposes.

Figure 3: Workflow of our layered-QAS. At generation

(i+1)

, the algorithm expands the PQC with a layer

\mathbf{L}_{i+1}(\theta_{i+1})

, chosen from single‑qubit, entangling, or pruning layers. Single‑qubit and entangling layers add new parametrised gates (green), while pruning layers remove gates (red). Several candidate layers are evaluated, and the most effective circuit extension is retained.

Figure 4: Layers tested in our layered-QAS. The number of parameters for each layer is shown at the top right. Special architectures include the pruning layer, which deletes gates, and architecture 3, which circularly applies CNOTs on coordinate registers.

	ModelNet10					ModelNet40
	NPQ	NPC	Tr.	Va.	Ts.	NPQ	NPC	Tr.	Va.	Ts.
Layered 10 [Ours]	$100$	$270$	$\mathbf{92\%}$	$\underline{92}\%$	$\underline{84}\%$	$105$	$1080$	$59\%$	$\underline{56}\%$	$\underline{54}\%$
Layered 20 [Ours]	$279$	$270$	$\mathbf{92\%}$	$\mathbf{93\%}$	$\mathbf{85\%}$	$279$	$1080$	$\mathbf{61}\%$	$\mathbf{59}\%$	$\mathbf{55}\%$
sQCNN-3D [ 1 ]	$48$	$650$	$83\%$	$79\%$	$72\%$	$48$	$2600$	$47\%$	$45\%$	$41\%$
Vanilla CNN	$0$	$590$	${90}\%$	$88\%$	$82\%$	$0$	$1580$	$\underline{60}\%$	$55\%$	$54\%$

Table 1: Resources and performance comparison for the models. Keys: “NP(Q/C)” =Number of parameters in the quantum or classical backbones; “Tr./Va./Ts.” =Train/Validation/Test top-1 accuracy; The subscript in our layered approach refers to the number of layered-QAS iterations. Our proposed approach consistently outperforms the baselines in all settings.

Panel of Figure 5 — Figure 5: Convergence of the search algorithms on the reduced ModelNet10 (top) and ModelNet40 (bottom) datasets for models with learnable (left) and frozen (right) linear layers. Shown are the accuracies of the top three candidates for the layered and evolutionary searches, and the single candidate for the local search. Layered search steadily improves accuracy by increasing PQC expressivity. Evolutionary search converges faster with fine‑tuning, while the version without fine‑tuning can occasionally find better models but struggles when the linear layer is frozen. Local search exhibits a step‑like behavior, as the model is updated only when an improvement is found.

	ModelNet10					ModelNet40
	NPQ(L/F)	NPC	Tr.(L/F)	Va.(L/F)	Ts.(L/F)	NPQ(L/F)	NPC	Tr.(L/F)	Va.(L/F)	Ts.(L/F)
Layered 10 [Ours]	$100/207$	$270$	$\mathbf{92/78\%}$	$\underline{92}/\mathbf{78}\%$	$\underline{84}/\underline{67}\%$	$105/164$	$1080$	$59/\mathbf{27}\%$	$56/\mathbf{26}\%$	$\underline{54}/\mathbf{21}\%$
Layered 20 [Ours]	$279/282$	$270$	$\mathbf{92/78\%}$	$\mathbf{93/78\%}$	$\mathbf{85/69\%}$	$279/220$	$1080$	$\mathbf{61}/\underline{22}\%$	$\mathbf{59}/\underline{22}\%$	$\mathbf{55}/\underline{16}\%$
Local search	$163/165$	$270$	$\underline{90}/\underline{75}\%$	$89/\underline{75}\%$	$83/65\%$	$166/166$	$1080$	$\underline{60}/20\%$	$\underline{58}/18\%$	$54/13\%$
Evolutionary [ 16 , 33 ]	$162/163$	$270$	$88/68\%$	$85/65\%$	$80/53\%$	$162/168$	$1080$	$59/18\%$	$57/17\%$	$52/12\%$
Evolutionary ft. [ 16 , 33 ]	$163/176$	$270$	$89/69\%$	$85/68\%$	$81/56\%$	$163/166$	$1080$	$59/17\%$	$58/16\%$	$53/11\%$

Table 2: Ablation study on the choice of QAS: “NP(Q/C)” =Number of parameters in the quantum or classical backbones; “Tr./Va./Ts.” =Train/Validation/Test top-1 accuracy; “L/F” =Search models with learnable and frozen linear layers; “ft.”=Fine tuning. NPC is

0

for models with frozen linear layers. The subscript in the layered model refers to the number of generations.

Figure 6: Performance of the layered search for different pruning thresholds

t

and voxelisation granularity

k

on the reduced ModelNet10. Pruning reduces the number of parameters without sacrificing accuracy. Increasing

k

increases the prediction accuracy.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 7 — Figure 7: Bloch sphere representation of a qubit. The qubit $\ket{\psi}=\alpha\ket{0}+\beta\ket{1}$ can be expressed as a unit vector in $\mathbb{R}^{3}$ . Two angles $\gamma\in[0,\pi]$ and $\phi\in[0,2\pi]$ fully describe the qubit in the basis spanned by the vectors $\vec{x},\vec{y}$ and $\vec{z}$ .

Figure 8: Voxelised point clouds with different granularity

k

. Brighter voxels are denser voxels, i.e . voxels containing most of the points, and darker voxels are less dense voxels. Higher values of

k

allow for capturing more details on the shape to classify. We use

k=3

in our experiments.

Figure 9: Candidate layers used in the evolutionary search. Note that the gate types and qubits, including controlled and target qubits of controlled gates, are chosen randomly in each layer.

Panel of Figure 10 — Figure 10: Convergence plot of the search algorithms on the full ModelNet10 dataset. Shown are the accuracies of the three best candidate models for each search strategy. The Layered search improves the prediction accuracy by enhancing the PQC expressivity. The evolutionary search, even with fine-tuning, shows only slight improvement over the generations.

Panel of Figure 11 — Figure 11: Benchmark performance of found models. Reported are top-1 accuracies on the respective sets.

Panel of Figure 12 — Figure 12: Confusion matrices on the test set of the ModelNet10 dataset. Models with learnable linear layers are more accurate than those with frozen linear layers. All models consistently misclassify similar shapes like desk & table , bathtub & bed , or toilet & chair . The latter pair is particularly difficult for models with frozen linear layers. Note that learnable and frozen linear layers do not apply to the vanilla CNN and sQCNN-3D models on the last row. “ft.” stands for fine tuning.

Explore similar work

Evolving Hybrid Quantum-Classical Architectures for Image Classification
Oct 2, 2026Devroop Kar, Daniel Krutz, Travis DesellHybrid Quantum-Classical ML Variational Quantum Circuits
Rochester Institute of Technology Rochester, NY 14623, USA
Investigating Quantum-Embedded Transformers on Classical Datasets for Cross-Modality Classification
Aug 7, 2026Hao-Yuan ChenHybrid Quantum-Classical ML Variational Quantum Circuits
Computer Science, University of London
Beyond Barren Plateaus: A Scalable Quantum Convolutional Architecture for High-Fidelity Image Classification
Mar 11, 2026Radhakrishnan DelhibabuQuantum Convolutional Neural Networks Quantum Circuit Trainability
School of Computing Science and Engineering (SCOPE), Vellore Institute of Technology, Vellore, 632014, India.

Oct 2, 2026J/K move · Enter open · S save

Evolving Hybrid Quantum-Classical Architectures for Image Classification

Devroop Kar, Daniel Krutz, Travis Desell

Rochester Institute of Technology Rochester, NY 14623, USA

Hybrid Quantum-Classical ML Variational Quantum Circuits Image Classification

Open paper AskPDF arXiv

Hybrid quantum classical neural networks integrate parameterized quantum circuits (PQCs) with established deep learning architectures, but their performance depends strongly on the choice of quantum circuit architecture, a choice that remains largely manual. Most existing approaches rely on hand-designed or fixed circuit ansätze, requiring circuit structure, gate composition, and qubit connectivity to be specified in advance with no guarantee that they suit the task. This limitation is especially acute in image classification, where quantum circuits must transform features extracted by classical networks while remaining compact enough for practical training, requirements that generic, task-agnostic ansätze are unlikely to satisfy simultaneously. We extend EXAQC, an evolutionary framework for automated quantum circuit discovery, to image classification. EXAQC evolves PQCs as intermediate processing modules while retaining classical feature-extraction and prediction layers. On MNIST, Fashion-MNIST, and CIFAR-10, EXAQC achieves 98.42%, 90.62%, and 85.47% accuracy, respectively, while using comparable gate counts to other quantum architecture-search methods. Against classical networks, evolved hybrid models maintain comparable accuracy with substantially fewer trainable parameters, reaching 85.68% on CIFAR-10 with over 25

\times

fewer parameters than a 10-layer CNN. Encoding choice also matters: rotation-based encodings (RX, RY, U3) outperform amplitude encoding by 22-25 points on CIFAR-10. These results demonstrate that automated circuit discovery yields compact quantum modules that can replace larger classical components in vision architectures while retaining competitive accuracy.

Abstract

Explore similar work

Evolving Hybrid Quantum-Classical Architectures for Image Classification

Investigating Quantum-Embedded Transformers on Classical Datasets for Cross-Modality Classification

Beyond Barren Plateaus: A Scalable Quantum Convolutional Architecture for High-Fidelity Image Classification