Diffusion models (DDPMs) have revolutionised generative modelling, surpassing GANs in images, advancing audio synthesis, and enabling de-novo protein design. Yet progress on time series lags behind early adversarial work. Recent studies highlight the benefits of spectral biases - FourierFlow and frequency-domain DDPMs.
In parallel, diffusion in the wavelet domain has emerged for images, offering a multi-resolution view well-suited to non-stationary signals. Wavelets capture localised, scale-dependent features, making them attractive for domains from finance to climate and biomedical data such as ECGs.
This project proposes the first DDPM framework operating directly in the wavelet domain for time series, aiming to improve generalisation, interpretability, and robustness across diverse sequential tasks.
- Artificial Intelligence and Signal and Image Processing, Statistics
- ETH Zurich (ETHZ), Master Thesis
