Numerical approximation of stochastic differential equations for sampling and generative modeling
Dnr:
NAISS 2026/4-1060
Type:
NAISS Small
Principal Investigator:
Akash Sharma
Affiliation:
Chalmers tekniska högskola
Start Date:
2026-06-04
End Date:
2027-07-01
Primary Classification:
10105: Computational Mathematics
Allocation
- Arrhenius Disk at NAISS: 259 GiB
- Arrhenius GPU at NAISS: 150 GPU-h/month
- Arrhenius Flash at NAISS: 50 GiB
Abstract
We are working on stochastic differential equations driven diffusion models. We are proposing diffusion models interlaced with jump process to accelerate their sampling process and new sampling and optimization methods based on stochastic differential equations. Our models have shown improved performance on toy data sets, we aim to deploy them on real-world data set which requires the use of GPU.