Generative deep learning models
The activity investigates the development of generative models to be used for data generation and as priors for inverse problems.
We develop generative adversarial networks (GAN) for data that are defined on regular grids (e.g., images), as well as graphs (e.g., point clouds). For these latter data, classical neural networks fail to capture effective features. Neural networks on graphs are expected to be able to learn both the graph describing data dependencies, as well as the related features.
We have successfully trained the first graph-GAN with application to point clouds, and we are also applying similar techniques to biological data.
We are also developing diffusion models, with a specific focus on reducing the complexity of the training process.
- PE7_7 Signal processing
- Unsupervised learning
- Artificial neural networks