Machine Learning in Neural Engineering

Department Empirical Inference
Max Planck Institute for Intelligent Systems

Machine Learning for Cognitive Neuroscience

In this project, we aim to advance the state-of-the-art of machine learning algorithms in cognitive neuroscience. Besides the development of algorithms for brain-state decoding, we are particularly interested in developing causal inference methods that provide novel insights into the neural basis of cognition.

More details on this project can be found in the following selected publications:

Transfer Learning in Brain-Computer Interfaces (IEEE Computational Intelligence Magazine, 2016) We provide an overview over transfer learning techniques in BCI research and present a novel decoding algorithm that treats decoders as samples from a normal distribution. This enables us to use transfer knowledge between subjects and/or sessions in a statistically rigorous fashion.

Identification of causal relations in neuroimaging data with latent confounders: An instrumental variable approach (NeuroImage, 2015) We provide an algorithm to test for causal relations between two brain processes, even in the presence of latent confounders.

Causal interpretation rules for encoding and decoding models in neuroimaging (NeuroImage, 2015) We provide an exhaustive set of rules which causal conclusions are supported and which ones are not warranted by empirical data in neuroimaging studies.

How to Test the Quality of Reconstructed Sources in Independent Component Analysis (ICA) of EEG/MEG Data (PRNI 2013) A simple method, based on EEG/MEG volume conduction models, to quantify the neurophysiological plausibility of reconstructed independent components.

Critical issues in state-of-the-art brain-computer interface signal processing (Journal of Neural Engineering, 2011) A review article that discusses the current state-of-the-art and most promising research directions in signal processing and feature extraction for BCIs.

Multitask Learning for Brain-Computer Interfaces (AISTATS, 2010) Learning a prior distribution over the feature space from previously recorded data can be used to compute a posterior classifier for new subjects that substantially enhances decoding accuracy and reduces calibration time relative to using subject-specific data only.

Beamforming in Noninvasive Brain-Computer Interfaces (IEEE Transactions on Biomedical Engineering, 2009) In noisy recordings, prior knowledge on the anatomical location of relevant brain areas can be used to learn spatial filters that substantially outperform competing methods in terms of decoding accuracy and are suitable for real-time feedback.

Multi-class Common Spatial Patterns and Information Theoretic Feature Extraction (IEEE Transactions on Biomedical Engineering, 2008) The Common Spatial Patterns (CSP) algorithm is shown to be optimal in terms of maximizing an (approximation of) mutual information of extracted features and class-labels. This insight is used to extend CSP to multi-class paradigms in an optimal fashion.