Understanding Score-Based Generative Models
Project: 00447
Diffusion models provide state-of-the-art tools to generate samples from complex distributions like images. However, there is often no analytic expression for the characterization of such distributions, which makes it hard to analyze such models. In this project, our goal is to shed light on the effect of the different components of the diffusion models. To do so, we use an analytic stochastic process to generate one-dimensional signals. We use such samples to train the diffusion model. Since we have the underlying generation process, we can benchmark the diffusion model and investigate whether the theoretical assumptions are satisfied in practice. We will also examine the effect of the neural network architecture in such models. The student needs to be familiar with deep learning and stochastic processes. A good level of PyTorch programming is necessary.
The image is taken from arXiv:2011.13456 (2020).
Related papers:
1) https://proceedings.neurips.cc/paper_files/paper/2019/hash/3001ef257407d5a371a96dcd947c7d93-Abstract.html
2) https://openreview.net/forum?id=PxTIG12RRHS&utm_campaign=NLP%20News&utm_medium=email&utm_source=Revue%20newsletter
- Supervisors
- Mehrsa Pourya, mehrsa.pourya@epfl.ch
- Pakshal Bohra, pakshal.bohra@epfl.ch
- Stanislas Ducotterd, stanislas.ducotterd@epfl.ch