This joint proposal gathers the GPU-intensive research of the Division of Statistics and Machine Learning (STIMA) at Linköping University, with Professor Fredrik Lindsten as PI. It comprises eight projects led by senior researchers, postdoctoral researchers and PhD students.
The projects cover (i) probabilistic weather forecasting and machine-learning-based data assimilation at kilometre scale; (ii) representation and transfer learning on raw LC-HRMS metabolomics data for forensic medicine; (iii) learned simulators and generative models for Lagrangian fluid dynamics and particle systems; (iv) estimation for branching stochastic processes in evolutionary biology; (v) deep anomaly detection and sensing in distributed MIMO wireless systems; (vi) generative models for synthetic medical images; (vii) scalable one-step generative diffusion models; and (viii) systematic evaluation of the energy and carbon cost of resource-efficient machine learning.
All projects involve training and evaluating models on large inputs, such as high-resolution atmospheric fields, mass spectra with hundreds of millions of measurements per sample, 3D medical images, large particle systems and multi-antenna channel data, which require high-memory GPUs and multi-GPU training. We request 5,000 GPU-hours/month on Arrhenius GPU, 5,000 core-hours/month on Arrhenius CPU and about 32 TB of storage. Results will be published at leading machine learning venues (NeurIPS, ICML, ICLR, AISTATS) and in domain journals.