PhD Research Seminar: Modern Methods of Machine Learning in the Interpretation of Electrocorticographic Activity of the Brain

Speaker: Artur Petrosyan, second-year PhD student, School of Data Analysis and Artificial Intelligence, Faculty of Computer Science
Where: https://zoom.us/j/442268392  
When: April 27, 19.40–21:00 

Neural interfaces are systems that provide exchange of information between the nervous system and an external device. An accessible technique that is now gaining popularity in the field of brain-computer interface (BCI) research is electrocorticography, a minimally invasive neuroimaging method in which electrical signal is sampled from electrodes placed either epidurally (on the surface of the dura mater) or subdurally (under the dura mater, on the surface of the brain). One way to use an electrocorticogram (ECoG) is to decode the kinematic properties of finger movement and to improve the accuracy of the control of the hand prostheses through the signals sent by the user's brain. Improving the quality of ECoG signal decoding is a key element in building a complete system of neurointerfaces. In this talk, we will focus on the task of decoding ECoG signals into an acceleration signal. There are already many neural network architectures whose structure is fully or partially motivated by prior knowledge in signal processing and neurobiology. Interpretation of weights of such simple architectures is usually reduced analysing spatial and temporal patterns. For spatial patterns, the weights considered by Hauf are usually analyzed. For temporal patterns, the standard Fast Fourier transformation (FFT) is used. However, both of these ways of interpreting patterns are incomplete. In particular, the temporal weights can take into account the non-white noise in the data, and, therefore, applying FFT will show a distorted temporal pattern. To make sure that our reasoning is correct, we made simulations with synthetic data. We also demonstrate the performance of such neural network of BCI competition to verify, that it has reached at least as good performance as the usual pipeline based on a linear regression model. Finally, we show patterns  learned by neural network on new data collected in our laboratory with known coordinates of each electrode.