September 12, 2017 15:00 - 16:00
BSI East Building 1F Seminar Room
Intelligent neurotechnology based on electroencephalographic (EEG) recordings has a broad application potential in healthcare and non-medical scenarios. Examples include EEG-informed medical assistive systems, intention-aware robotics utilizing online EEG monitoring, and automated neurological EEG diagnostics. Such applications will however require methods to more reliably decode information from EEG signals than possible today, and could also profit from optimized EEG acquisition.
Our work shows that high-gamma band (HGB, above 50 Hz) brain activity, a highly informative brain signal component, is detectable with optimized EEG methods. Previously it was thought that intracranial electrodes are required for this. To optimize information extraction from EEG signals in general, and HGB components in particular, we have adopted deep learning with convolutional neural networks to EEG analysis. I describe the rationale of our network architectures, which are global in space and local/hierarchical in time. Our experiments show that recent advances from deep learning research, for example the recently introduced exponential instead of rectified linear units, are important to reach high decoding accuracies. We also developed methods to visualize what the networks learned from the EEG data, both using correlative and perturbation-based (causal) techniques. Our deep learning methods and visualization tools are available in the “braindecode” open-source toolbox on GitHub.
We compared the performance of our ConNets to various widely used methods on a broad range of decoding problems, including decoding of movement parameter from EEG spectral power changes, cognitive control signals, information about pathological EEG states, as well as information about the correctness of actions, both in non-invasive as well as intracranial EEG. ConvNets robustly performed at least as good, or better, than the classical methods. ConvNets were always trained end-to-end, while the competitor methods often required a priori decisions about the features to be used. Visualization showed clear spatio-temporal patterns in the signals used by the ConvNets, up the high EEG gamma range. These visualizations thus allowed new insight into informative EEG features. In an online experiment, a continuously adapting version of our ConvNets allowed the control of a semi-autonomous robotic drinking assistant. There are many further techniques that have been already successfully been used in the deep learning computer vision field that my further boost performance and that we are currently adopting to EEG, including data augmentation, recurrent networks, ResNets, and hyperparameter optimization. Thus deep learning techniques, combined with optimized intracranial and non-invasive recording methods, may help advancing closed-loop neurotechnology to real-world applications, and may open up new possibilities in brain signal analysis and visualization.
- Open to Public
- Keiichi Kitajo [Keiichi Kitajo, Rhythm-based Brain Information Processing Unit, BTCC ]
Name: Keiichi Kitajo