Exzellenzcluster – Maschinelles Lernen für die Wissenschaft
Unsere Stipendiaten
Stipendiaten 2024
Die 5 AIMS-Stipendiaten in Tübingen. Von links nach rechts: Fifamé Edwige Akpoly, Pauline Ornela Megne Choudja, Stephen Kiilu, Berthine Nyunga Mpinda and Immaculate Wanjiru Kimani.
Forschungsprojekt: Explainable Medical Image Classification
Stipendiaten 2023
Die 5 AIMS-Stipendiaten in Tübingen. Von links nach rechts: Mendrika Rakotomanga, Albert Agisha, Ifeoma Nwabufo, Amel Abdelraheem, und Fenosoa Randrianjatovo
Machine learning and statistics for climate networks In this project, we investigate novel methods to assess the uncertainty in climate network construction and estimation procedures. The goal will be to create uncertainty estimates, not by network subsampling, but by bootstrapping the underlying time series directly. Different approaches will be compared, both on a theoretical and implementation level and connected to background knowledge in climate science.
Unsupervised learning in medical data science The project will look at unsupervised learning methods recently developed in the lab to embed medical images via contrastive learning. It will explore the resulting embedding, study which data transformations are important and whether it can be related to disease/demographic factors.
Bayesian inference of gravitational waves with machine learning This project is concerned with the evaluation and development of machine learning methods for gravitational wave inference. One focus is the analysis of current methods with regard to possible biases and their impact on large-scale analyses. In this context, the project also aims to develop methods to mitigate possible inaccuracies.
About random strategies in (non-stationary) reinforcement learning Reinforcement learning (RL) algorithms have so far been developed mainly for steady-state environments and have difficulty adapting when the system dynamics or the reward function changes. Posterior sampling and Thompson sampling were identified early on as efficient strategies in RL, in part due to their randomised strategies. While some algorithms such as UCRL have recently been adapted to non-stationary environments, no randomised strategy has been proposed yet. In this project, we aim to propose a randomised -and more practical- algorithm that builds on posterior sampling and is capable of achieving sublinear regret. We will start by studying a (possibly context-dependent) bandit problem and then extend our findings to more complex RL models.
| Wir trauern um Bolaji, die am 19. September 2023 auf tragische Weise bei einem Autounfall ums Leben kam. Unsere Gedanken sind bei der Familie und den Freunden von Bolaji. Wir erinnern uns mit Freude an unsere Zeit mit ihr in Tübingen.|
