| name | qdiffusion-time-series |
| description | Quantum generative diffusion model for time series synthesis (QDiffusion-TS). First quantum diffusion model validated on real quantum hardware (IQM processor). Replaces feed-forward components with quantum neural networks, reducing parameters by ~3 orders of magnitude. Reduces Wasserstein distance by ~44% vs classical. Use when: quantum generative modeling, time series synthesis, quantum-enhanced data augmentation, parameter-efficient diffusion models. |
Core Methodology
QDiffusion-TS Architecture
Extends classical diffusion architecture by replacing feed-forward components within the denoising transformer with Quantum Neural Networks (QNNs), yielding a hybrid quantum transformer.
Key Innovation: Parameter Reduction
- Reduces trainable parameters in each replaced component by nearly 3 orders of magnitude
- Maintains or exceeds classical performance with substantially fewer parameters
Performance Metrics
- Wasserstein distance: ~44% reduction vs classical counterpart
- Downstream forecasting: up to 71% improvement in RMSE with synthetic data augmentation
- Validated on financial time series (Apple, Amazon) on IQM quantum processor
Pipeline
Real Time Series → Diffusion Process (noisy)
→ Denoising Transformer with QNN layers
→ Synthetic Time Series
→ Downstream Task (e.g., forecasting)
Quantum-Classical Hybrid Design
- Quantum layers handle the core denoising transformation
- Classical layers handle preprocessing and postprocessing
- Joint training of quantum and classical components
Data Augmentation Strategy
Generated synthetic data is used to augment real training data, improving downstream model performance significantly (up to 71% RMSE improvement).
Implementation Notes
- Requires access to quantum hardware (IQM processor or similar)
- Hybrid architecture allows training with classical simulation, then deployment on quantum hardware
- Particularly effective for small datasets where quantum expressivity provides advantage
Activation
quantum diffusion, quantum generative model, time series synthesis, QDiffusion, quantum data augmentation, hybrid quantum transformer, quantum ML time series
