| name | quantum-gan-benchmark-controlled-evaluation |
| description | Controlled benchmark methodology for quantum vs classical generators in medical imaging with matched parameters (arXiv:2606.18970) |
| category | quantum-benchmarking |
Controlled Benchmark of Quantum Generative Augmentation
Methodology from arXiv:2606.18970 (June 2026). Rigorous framework for evaluating quantum generative models in medical imaging.
Core Pattern
Isolate quantum generator contribution through controlled evaluation:
- Match parameter budgets between quantum and classical generators (1648 vs 1632 params)
- Multiple random seeds with paired significance testing
- Multiple comparison correction
- Intraset diversity and latent-distribution analyses
- Evaluate across labeled data fractions (5% to 100%)
Key Findings
- No augmentation variant significantly outperforms real-data-only training
- Quantum and classical generators are statistically indistinguishable
- Low-data benefit acts as regularization, not faithful data expansion
- Synthetic samples are off-distribution and severely mode-collapsed where data is scarce
- Framework released as testbed for rigorous quantum generative evaluation
Implementation Steps
- Encode images into KL-regularized latent space
- Train conditional Wasserstein GAN with gradient penalty
- Compare quantum vs classical generator with matched parameters
- Evaluate across data fractions (5%-100%) with 8+ random seeds
- Apply paired significance testing with multiple-comparison correction
- Analyze intraset diversity and latent distributions
- Characterize whether benefits are regularization or faithful expansion
When to Use
- Evaluating quantum generative models for medical imaging
- Any claim of quantum advantage in generative modeling
- Need for rigorous, controlled benchmarking protocols
- Avoiding false positives from parameter budget mismatches
References
- arXiv: 2606.18970v2
- Authors: Syed Mujtaba Haider, Silvia Figini
