Layered Quantum Architecture Search for 3D Point Cloud Classification
Natacha Kuete Meli Jovita Lukasik Vladislav Golyanik Michael Moeller · 2026
Read Layered Quantum Architecture Search for 3D Point Cloud Classification by Natacha Kuete Meli Jovita Lukasik Vladislav Golyanik Michael Moeller (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-20024.txt.
Introduction
Kuete Meli and co-authors grow quantum circuits layer by layer to classify 3D point clouds. Parametrised quantum circuits are expressive but lack the built in layer types of classical networks. Their layered search borrows from network morphism, adapting the circuit as it grows. Earlier work used such circuits only as feature extractors. Here the quantum circuit is the main model. Simulations show the method softens the barren plateau problem. It beat local and evolutionary search baselines, and set the best ModelNet scores among circuit methods.
Catalog record
| CREATOR | Natacha Kuete Meli Jovita Lukasik Vladislav Golyanik Michael Moeller |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 56456 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.20024
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Layered Quantum Architecture Search for 3D Point Cloud Classification Natacha Kuete Meli Jovita Lukasik Vladislav Golyanik Michael Moeller 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 1 1 1 project page: https://4dqv.mpi-inf.mpg.de/LQAS/ . 1 Introduction Quantum Machine Learning (QML) 3 ; 5 ; 26 , through shallow-depth Parametrised Quantum Circuits (PQCs), is anticipated to pave the way for utility-scale quantum computing. PQCs operate in high-dimensional Hilbert spaces with comparably few parameters, leveraging superposition and entanglement to extract features beyond classical neural networks. However, designing effective PQC architectures for task-oriented feature extraction remains challenging. While classical models benefit from well-established architectural layers and inductive biases, PQCs lack such standardised building blocks. In addition, the same quantum properties that enable high expressiveness can also cause barren plateaus 30 , where flattened loss landscapes hinder optimisation. Furthermore, PQCs are largely composed of linear transformations, missing the non-linear activations that underpin classical deep networks’ representational power. In light of this, maximising PQC architectural design is crucial to ensure discriminative feature extraction while keeping the PQCs trainable and robust. 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. Quantum Architecture Search (QAS) offers a systematic way to address these design challenges by automatically exploring the space of PQC structures 59 ; 32 ; 34 . By replacing manual trial-and-error with guided search, QAS can discover expressive circuits for task-specific feature extraction, improve trainability, and mitigate barren plateaus. Existing QAS policies, however, face several limitations. Super-circuit and weight-sharing strategies 10 ; 47 not only search in very large spaces that are computationally demanding, but also often converge to suboptimal architectures, as candidate models can adversarially update the super-circuit parameters. Evolutionary and RL-based methods may struggle with scalability or training instability due to sensitivity to hyperparameters and reward design 39 , while differentiable approaches 51 are prone to getting trapped in local optima. Moreover, none of these QAS methods account for PQC-specific considerations such as strategic training 46 ; 11 and careful initialisation 15 ; 22 ; 48 , which have been shown to mitigate barren plateaus, improve generalisation, and enhance feature learning. In response to these limitations, we introduce layered-QAS . Building on ideas from the Lamarckian-based LEMONADE 12 , our policy adds new layers to a pre‑trained circuit and retains only those that provide the largest performance improvement. To further increase efficiency, we incorporate a pruning mechanism that removes gates operating near the identity, reducing unnecessary complexity without sacrificing expressivity. Together, these components make layered‑QAS a robust framework for discovering high‑performing PQCs. We target 3D point cloud classification as a demanding yet structured exemplary application to evaluate our approach. The task involves assigning labels to objects represented by unordered sets of points, with applications in autonomous driving, robotics, and semantic segmentation 56 ; 41 ; 37 . Point clouds vary in density, resolution, and shape, while modern sensors produce increasingly large datasets. These characteristics make point cloud classification an ideal benchmark for QML, as it requires models to reason about spatial relationships in irregular, high-dimensional data. In this context, sQCNN-3D 1 applies quanvolutional filters to point cloud patches but delegates classification to a classical fully connected network—leaving the quantum model underutilised. In contrast, we pursue a fully quantum classification pipeline, minimising classical post-processing and improving feature extraction by the quantum circuit through architecture search. Applied to 3D multi-class point cloud classification, our pipeline enables near end-to-end quantum classification with only minimal classical components. An overview of our 3D classification framework is provided in Figure 1 . To summarise, our main contributions are as follows: • A new layered-QAS policy for discovering improved and task-adapted PQC designs, avoiding manual trial-and-error approaches (Sec. 3 ; Sec. 3.2 ); • A new framework for 3D point cloud classification based on PQCs and amplitude encoding of 3D data, leveraging layered-QAS for architectural design (Sec. 3.1 ). We experiment on the ModelNet datasets 52 using a quantum computer simulator and show that our models outperform the existing sQCNN-3D quantum baseline 1 and are even competitive with a classical baseline of similar expressivity that we design. Our search strategies can find parameter-efficient PQC architectures. Moreover, our layered search achieves better performance than the evolutionary search and a QAS-adapted local search 49 , a baseline that incrementally improves PQC architectures by making small localised changes to the PQC and retains the changes that enhance performance. 2 Related Work 2.1 PQC Training and QAS Strategic PQC Training. Strategic training of PQCs is believed to dampen trainability issues. Skolik et al. 46 proposed a layer-wise training approach, which incrementally grows the circuit depth, freezes previously learned parameters, and optimises only the parameters of added layers. This layered training strategy not only speeds up the training but also experimentally proved to increase the generalisation error. Similarly, the method proposed by Duffy et al. 11 allows not only for the addition of new parametrised gates, but also feature-map encodings of the data to incrementally grow the circuits. Data re-uploading is proven to increase the expressivity of PQCs 43 . Both methods suggest that incrementally training PQCs is a promising approach to enhance PQCs’ expressivity without compromising their trainability. Note that a layered training approach was also used by Krahn et al. 24 for binary networks with quantum annealed gradients. Strategic PQC Initialisation. Random initialisation of PQCs was shown to be one cause of barren plateaus 35 ; 17 . Grant et al. 15 proposed a selective initialisation method that randomly selects only a subset of the initial parameter values and chooses the remaining ones so that the circuit is a sequence of shallow blocks that evaluate to the identity, limiting the circuit depth in the first parameter update. Wang et al. 48 proved that reducing the initial domain of each parameter inversely proportional to the square root of the circuit depth causes the magnitude of the cost gradient to decay at most polynomially with respect to the qubit count and the circuit depth. Kashif et al. 22 empirically showed that initialising parameters within smaller distribution ranges, with lower magnitudes, tends to perform better than using larger ranges