Machine Learning and Artificial Intelligence (MLI)
The MLI group – established at HITS in September 2022 – works on novel algorithms and models for data-efficient learning, geometric deep learning, and interpretability.
Data-efficient learning enables to take an existing model, e.g. that was trained on a big standard dataset, and adapt this model to a novel application. The dataset of the new domain that is used for fine-tuning the model then can be much smaller than it would have to be without pre-training. This reduces the time and resources needed for collecting training data and enables the use of machine learning approaches in application areas which so far could not benefit from this technology.
A particular research focus of the group is in developing data-efficient methods with improved generalization properties, which reduce the expected error of the fine-tuned model in practice. This is achieved by combining insights from learning theory with methods from convex and non-convex optimization.
Another research focus is on learning of interpretable representations, with the goal to make models interpretable and therefore enable a better understanding of the underlying principles, and to discover generative factors in data. Some of these methods enable to reconstruct causal relationships from observed data with exciting applications especially in the natural sciences. For this the group applies and extends methods from variational inference, statistical methods for independent component analysis, and graph-neural-networks.
We have several PhD and postdoc positions available. Please apply using our career website.