| name | cv-photonic-qnn-edge-ai |
| description | Continuous-variable photonic quantum neural networks for parameter-efficient edge AI medical imaging with room-temperature operation and extreme parameter reduction |
| tags | ["quantum","photonic","cv-qnn","edge-ai","medical-imaging","parameter-efficient","room-temperature","oral-cancer"] |
CV-Photonic QNN for Edge AI Medical Imaging
Paper Summary
Title: Parameter-Efficient Continuous-Variable Photonic Quantum Neural Networks for Edge Quantum AI: Demonstration in Oral Cancer Detection
arXiv: 2606.28252
Authors: Akshay Bhagwan Sonawane, Sophie Choe, Lakshman Tamil
Date: June 26, 2026
Core Innovation
Hybrid classical-CV quantum classifier achieving 100% test accuracy on oral cancer detection with only 18 trainable parameters (67% fewer than 55-parameter classical baseline).
Technical Architecture
Pipeline
- MobileNetV1 feature extractor (pretrained)
- PCA dimensionality reduction to 16 features
- CV-QNN classifier with:
- Displacement gates
- Interferometric gates
- Kerr nonlinearity gates
- Photonic backend (room temperature)
Simplified CV-QNN Layer
- Parameter reduction: 40-45% fewer parameters vs. standard CV-QNN (Killoran et al. 2019)
- Barren plateau mitigation: Dimensionality reduction + encoding restriction raises loss-gradient variance by ~58 orders of magnitude
- Width-dependent advantage:
- 2 qumodes: Full layer slightly better
- 4 qumodes: Simplified layer significantly better (44% fewer parameters)
Best Model Performance
- Architecture: 4-qumode simplified CV-QNN
- Parameters: 18 trainable
- Validation AUC: Highest among all models tested
- Test accuracy: 100% calibrated across all random seeds
- Parameter efficiency: 67% fewer than 55-parameter classical baseline
Key Advantages
Edge Deployment
- Room-temperature operation: No cryogenics required (unlike qubit-based QNNs)
- Smartphone-compatible: Lightweight enough for mobile inference
- Low-resource settings: Targets oral cancer screening in areas lacking specialized diagnostic tools
Quantum Advantage
- Parameter efficiency: Extreme compression (18 params vs. 55 classical)
- Expressivity: Compact representation of complex medical image distributions
- Trainability: Barren plateau mitigation enables practical optimization
Implementation Notes
CV-QNN vs. Qubit-Based QNN
| Feature | CV-Photonic QNN | Qubit-Based QNN |
|---|
| Temperature | Room temperature | Cryogenic (~15 mK) |
| Hardware | Photonic chips | Superconducting circuits |
| Scalability | Natural for optical systems | Limited by qubit count |
| Use case | Edge deployment | High-fidelity quantum computing |
Barren Plateau Mitigation
Two strategies combined:
- Dimensionality reduction: PCA to 16 features
- Encoding restriction: Limit input encoding complexity
Result: Gradient variance increased by ~58 orders of magnitude
Reproducibility
- Dataset: Smartphone oral cancer images (specific dataset not detailed in abstract)
- Classical baseline: 55-parameter model
- Quantum backend: Photonic (specific platform not specified)
- Training: Standard CV-QNN optimization with simplified layer
Potential Extensions
- Larger medical datasets: Breast cancer, skin cancer screening
- Real-time inference: Mobile app integration
- Federated learning: Privacy-preserving multi-site training
- Hybrid architectures: Combine CV-QNN with classical neural networks
Related Work
- Killoran et al. (2019): Standard CV-QNN layer (baseline for comparison)
- MobileNetV1: Lightweight feature extraction for mobile deployment
- Quantum reservoir computing: Alternative quantum ML approach for time series
When to Use
- Medical imaging with extreme parameter constraints
- Edge deployment requiring room-temperature quantum hardware
- Scenarios where quantum advantage in parameter efficiency outweighs absolute accuracy
- Low-resource healthcare settings needing smartphone-compatible AI
Limitations
- Small dataset (oral cancer detection specific)
- Photonic hardware platform not specified (reproducibility concern)
- No comparison with other quantum ML approaches (e.g., quantum reservoir computing)
- 100% accuracy may indicate overfitting on small dataset
