Section: New Results

Higher level functions

Participants : Frédéric Alexandre, Laurent Bougrain, Octave Boussaton, Axel Hutt, Baptiste Payan, Maxime Rio, Carolina Saavedra, Christian Weber.

Our activities concerned information analysis and interpretation and the design of numerical distributed and adaptive algorithms in interaction with biology and medical science. To better understand cortical signals, we choose a top-down approach for which data analysis techniques extract properties of underlying neural activity. To this end several unsupervised methods and supervised methods are investigated and integrated to extract features in measured brain signals. More specifically, we worked on Brain Computer Interfaces (BCI).

Detection of partial amplitude synchronization in multivariate data

To gain information on the interactions between neural structures, several electrodes may be implanted in cortical areas to measure Local Field Potentials. The developed method aims to extract time windows in which a subset of measured time series exhibit an amplitude synchronization in certain frequency bands [12] .

Brain-Computer Interface based on motor imagery to control a robotic arm in 3D

The interface we develop aims to control in 3D a Jaco robotic arm by Kinova, using the Graz Motor Imagery detection paradigm for two or three motor actions in an online situation. The interface is part of the OpenViBE software. The user can switch in different modes to control a specific part of the robotic device (arm, wrist, fingers). We plan to use five different motor imageries: right hand, left hand, foot, rest and both hands. The actions are not available all together for a specific control. The interface is already done. More experiments will be done to adjust the classifier.

Reinforcement learning to better control a robotic arm

The approach we proposed in Cobras is innovative. Many studies attempts improve the recognition rate of a BCI order with new methods for treatment of signal. These studies are placed upstream of the BCI to facilitate the retrieval of information in the signal. However, the signal to noise ratio is so low that the improvements are limited. Rather than improving signal processing upstream, we wanted to improve the recognition rate by adding information in the controlled system. Thus, we placed downstream and added, as inputs of our control system, mechanical data concerning robotic arm. Initially, we validated the possibility of finding -using an inverse algorithm of reinforcement learning- the policy of the expert from a set of trajectories followed in a maze. We defined then a scenario to achieve different trajectories with the robotic arm to reach several buttons. In a third step, we used this algorithm on a maze-type problem but for which we have completed the state vector with the classifier outputs. This study is ongoing.

Mutual influence of firing rates of corticomotoneuronal cells for learning a precision grip task

As a part of a Brain-Machine Interface, we define a model for learning and forecasting muscular activity, given sparse cortical activity in the form of action potential signals (spike trains). We have a collection of experiments in which a trained monkey performs a precision grip. More precisely, its neuronal activity is partially recorded from corticomotoneuron cells of the hand area (area 4) as the monkey clasps two levers between its index finger and thumb. The underlying model parameters are interpreted with respect to the physiological aspects, though the model itself is not bio-physical. The method used is based on a system of first degree linear equations involving the firing rate of the recorded neurons, two sets of thresholds associated to them, and the variation of the global neuronal activity. We build a module to translate the data in the form of spikes trains into the event structure of OpenViBE triggers which is more appropriate than signals. The enslavement of the clamp according to the order generated by OpenViBE was also done. These solutions can demonstrate the capabilities of our algorithms for decoding cortical signals in the task of handling.

Hysteresis thresholding for Wavelet denoising applied to P300 single-trial detection

Template-based analysis techniques are good candidates to robustly detect transient temporal graphic elements (e.g. event-related potential, k-complex, sleep spindles, vertex waves, spikes) in noisy and multi-sources electroencephalographic signals. More specifically, we studied the significant impact on a large dataset of wavelet denoisings to detect evoked potentials in a single-trial P300 speller. We applied the classical thresholds selection rules algorithms and compared them with the hysteresis algorithm by R. Ranta which combine the classical thresholds to detect blocks of significant wavelets coefficients based on the graph structure of the wavelet decomposition.