We hypothesize that meta-learning will be generally useful for machine learning on small data problems when multiple heterogeneous small data sets are available for pre-training. We aim to develop and evaluate such methods, motivated by the clinically relevant use case of brain signal analysis, using deep neural networks, where challenges of heterogeneous EEG data will specifically be addressed. We will develop techniques for information transfer across small EEG datasets from different setups, subjects, and tasks. Specifically, we will collect publicly available EEG data sets, and train a single large neural network across the whole collection by learning to align different topological layouts, and training a transformer model that can directly process the electrode coordinates. Moreover, we will pre-train neural networks on subsets of homogeneous EEG datasets and study how to exploit pre-trained neural networks to fine-tune for a new task.
(supervised by PI Ball)
Working with the EEG data and work on developing neural networks methods for EEG decoding.
(supervised by PI Hutter)
Approaches for meta-learning how to fine-tune large pre-trained models.
Institute of Medical Biometry and Statistics,
Faculty of Medicine and Medical Center –
University of Freiburg