In the recent time there are many developments in the field of the brain - computer interface (BCI).
Abstract
In the recent time there are many developments in the field of the brain – computer interface (BCI). This field is intended to provide solution to handicapped people that can’t use their hands to control computer. First, the person using BCI, trains it by concentrating on some specific mental task. Usually 2 different tasks are used in the training. BCI registers several EEG samples of each task. Ones the training is complete, the subject concentrate on one of this tasks again, and the system is required to “guess” what task the person is concentrating on.
The problem
The problem of classifying different tasks is not easy to solve. Many different techniques already developed, but none of these technique works perfect (with 0% of misclassification or close to that). The aim of this project is to propose another technique in attempt to obtain the lowest rate of misclassification.
The solution
Our method is based on the assumption that the “real brain signal” of each task is smooth, and is contained in each sensors channel. So we have attempted to reveal the “real impulse” by taking some linear combination of all sensor data, while optimizing on the smoothness of the result average. To obtain the weights with which we will sum all the channels we’ve developed the following optimization function:
F1(w)= -(norm2(w*(xavg_left – xavg_right)))2 + norm1(D(w*x))
We look for argmin(F1(W)).
Figure 1 – Original signal
Figure 2 – Restored signal
Tools
We used Matlab to define and solve optimization problem. We also used SVM (support vector machine) Matlab toolbox for classification purposes.
Reference: to be published…
Conclusions
Our method of preprocessing is proved to be useful in recovering the shape of smooth physiological signals. Although the algorithm has failed on original signals, the test case with synthetic signals has shown that the form of original signals can be restored. The failure on original signals can be explained by too noisy environment or lack of smooth form, which is required for our method. The quality of recovering depends on level of noise. However, even in the noisy environment (when the form of original signal cannot be view on channel’s average), our algorithm successfully restores the original signal. This algorithm can be successfully used as preprocessing in classification task, as well as in the blind source separation itself.
Acknowledgment
We are very thankful to our supervisor Michael Zibulevsky for his patients and guiding us throughout the project. We are also grateful to Johanan Erez for his help and support. We are also grateful to the Ollendorf Minerva Center Fund for supporting this project.