Tonio Ball

Translational Neurotechnology Lab
University Medical Center
Freiburg
Germany

Deep Learning for Neurotechnology
Lecture Abstract:

"Deep Learning (DL) based on artificial neural networks is a class of machine learning algorithms that has significantly improved the state-of-the-art in many domains, including computer vision and speech recognition tasks. DL enable the learning of hierarchies of increasingly abstract data representations, for example by training deep convolutional or recurrent neural networks (CNNs, RNNs) in an end-to end manner on the raw input signals. Today, there is a rapidly increasing interest in leveraging the advantages of DL for neuroscience as well as for healthcare applications. Here, I summarize the recent efforts in my lab to adopt DL techniques to brain-machine interfacing (BMI) tasks. I show how we designed a CCN architecture for online EEG decoding and successfully used it in the first adaptive deep-learning-based BCI for control of a robot assistant. Analysis of the internal EEG feature representations across the different CCN layers suggests that they learn hierarchical representations of temporal EEG features. Thus, the networks learned to use EEG spectral power changes not only in the alpha and beta, but also high gamma frequency ranges to infer the user’s commands. In other situations, where not spectral power but EEG phase was most informative, networks also learned to extract signal phase. In both cases, networks performed en par or better than traditional methods. Thus, DL for EEG decoding offers high accuracy, appears to flexibly adapt to the most informative features in a given task, and lends itself to adative online application. Beyond these first steps in online application of DL, there are many avenues for further research and development, such as through compilation of large high-quality public EEG data collections, advanced methods in network hyperparameter optimization, automatic architecture search and meta-learning, or through neuromophic hardware which may be particularly useful for implantable BMIs."