Layered Quantum Architecture Search for 3D Point Cloud Classification

1. Problem Statement

The paper addresses the challenge of designing effective Parametrised Quantum Circuits (PQCs) for 3D point cloud classification. While Quantum Machine Learning (QML) promises high expressiveness with few parameters, designing PQC architectures for specific tasks remains difficult due to:

• Lack of Inductive Biases: Unlike classical deep learning (which uses standard layers like convolution and attention), PQCs lack standardized architectural building blocks.
• Trainability Issues: PQCs are prone to barren plateaus, where the loss landscape becomes flat, making gradient-based optimization impossible.
• Limited Non-linearity: PQCs are primarily composed of linear transformations; they lack the non-linear activation functions that give classical networks their representational power, relying instead on quantum measurements for non-linearity.
• Inefficient Existing Approaches: Prior work on 3D point clouds (e.g., sQCNN-3D) often uses PQCs merely as feature extractors feeding into large classical networks, underutilizing the quantum model's potential.

2. Methodology: Layered Quantum Architecture Search (Layered-QAS)

The authors propose Layered-QAS, a strategy inspired by classical network morphism and Lamarckian evolution (specifically the LEMONADE algorithm). Instead of searching a massive space of random circuits, the method progressively grows and adapts a PQC.

A. Data Encoding (Amplitude Encoding)

• Input: 3D point clouds are voxelized into a grid of size $2^k \times 2^k \times 2^k$.
• Encoding: The density of points in each voxel is calculated and normalized to form a probability distribution. This is encoded into a quantum state vector $|\psi\rangle$ using amplitude encoding.
• Qubit Count: For a granularity $k$, the system uses $3k$ qubits (encoding $x, y, z$ coordinates). The paper uses $k=3$ (9 qubits).

B. The Layered-QAS Algorithm

The search process is iterative (generation $i$ to $i+1$):

1. Initialization: Start with a trivial identity circuit ($U_0 = I$) containing only data encoding and measurement.
2. Candidate Generation: In each step, a new layer $L_{i+1}$ is added to the existing circuit $U_i$. The layer type alternates cyclically among three types:
   • Single-qubit layers: Apply Pauli rotations ($R_X, R_Y, R_Z$) to individual qubits.
   • Entangling layers: Apply controlled rotations (e.g., $CR_X$) between qubit pairs to introduce correlations.
   • Pruning layers: Remove gates where the parameter magnitude is below a threshold (e.g., $\pi/10$), reducing complexity without sacrificing expressivity.
3. Evaluation: Multiple candidate circuits (different layer configurations) are trained for a few epochs.
4. Selection: The candidate with the highest validation accuracy is selected as the new base circuit $U_{i+1}$. Crucially, parameters from previous layers are retained and further optimized (warm-starting), rather than re-initialized.

C. Classification Pipeline

• Measurement: Qubits are measured in the Pauli X, Y, and Z bases, yielding $3 \times 3k$ features.
• Classifier: These features are fed into a classical linear layer (logits) followed by a Softmax and Cross-Entropy loss.
• Flexibility: The authors test both trainable and frozen (random Gaussian projection) linear layers to prove that the PQC itself learns discriminative features.

3. Key Contributions

1. Layered-QAS Policy: A novel, progressive architecture search method that avoids manual trial-and-error. It combines informed initialization, layer-wise training, and pruning to discover task-adapted PQC designs.
2. End-to-End Quantum Pipeline: A framework for 3D point cloud classification where the PQC is the primary feature extractor and classifier, minimizing classical post-processing.
3. Mitigation of Barren Plateaus: By using a progressive growth strategy (adding layers to a pre-trained circuit) and strategic initialization, the method mitigates barren plateaus and improves trainability.
4. Pruning Mechanism: A specific pruning strategy that removes near-identity gates to reduce circuit depth and parameter count without significant accuracy loss.

4. Experimental Results

The method was evaluated on the ModelNet10 and ModelNet40 datasets using a noise-free quantum simulator.

• Performance vs. Baselines:

  • vs. sQCNN-3D: The proposed Layered-QAS significantly outperformed the previous state-of-the-art quantum method (sQCNN-3D) on both datasets. For example, on ModelNet10, Layered-QAS achieved 85% test accuracy compared to sQCNN-3D's 72%.
  • vs. Classical Baselines: The quantum models achieved competitive accuracy with a "Vanilla CNN" baseline of similar expressivity but with far fewer parameters (e.g., 270 quantum parameters vs. 590 classical parameters for ModelNet10).
  • Frozen Linear Layer: Even with a frozen (random) linear layer, the PQC alone achieved ~78% accuracy on ModelNet10, demonstrating that the quantum circuit learned highly discriminative features. In contrast, sQCNN-3D with a frozen MLP dropped to ~17%.

• Comparison with Other QAS Methods:

  • Layered-QAS outperformed Evolutionary Search (using super-circuits) and Local Search.
  • Evolutionary search struggled with scalability and weight-sharing conflicts on full datasets, often converging to suboptimal architectures. Layered-QAS showed steady, monotonic improvement in accuracy as layers were added.

• Ablation Studies:

  • Pruning: Pruning gates with small parameters reduced the gate count by ~30% without hurting accuracy.
  • Granularity: Increasing voxel granularity ($k$) improved accuracy, but $k=3$ offered the best trade-off between detail and computational cost.

5. Significance and Conclusion

• Task-Specific Quantum Design: The paper demonstrates that automated architecture search is essential for unlocking the potential of PQCs in complex, structured domains like 3D vision.
• Efficiency: The method proves that quantum models can achieve high performance with extremely low parameter counts, addressing the "parameter efficiency" argument for QML.
• Trainability: The layered growth strategy effectively bypasses the barren plateau problem, offering a practical path to training deeper quantum circuits.
• Future Outlook: While the current models rely on linear PQC operations, the success of this architecture search suggests that future designs incorporating intermediate non-linearities (via measurement) or binary encodings could further bridge the gap between quantum and classical performance.

In summary, this work establishes Layered-QAS as a robust framework for designing quantum neural networks, successfully applying it to a challenging 3D classification task and outperforming existing quantum and classical baselines in terms of accuracy-to-parameter ratio.