Meeting ID: 97860059707
Password: ****** (Click the link below to find out more)
Session Link: Less is More
Brain computer interfaces (BCI) are amongst the exhilarating applications of computational neuroscience and have been increasingly the focus of research around the world. Currently, abundant effort in BCI research is devoted to offline analysis of recorded data to achieve higher accuracies in decoding [1]. Although this has led to the development of new methods and algorithms, the problem of online decoding of subject intentions remains a challenging one [2]. One of the restrictive steps of BCI design is the need for complex preprocessing steps required to extract features from the recorded signals to be used by the classifiers to distinguish intentions of the subject [1]. The other hindering factor is the variability of the recorded signals. EEG recordings use between 20-128 electrodes at sampling rates of 250 up to 1KHz for BCI applications. This data is recorded from the whole brain and in/between subject(s) variability intensifies the problem even further. This problem is currently mitigated through manual and careful feature engineering steps and tweaking of classifier parameters.
We are proposing to reduce the complexity of the architecture by 1)using only raw recorded signals with no preprocessing, 2)reducing the number of channels used for classification and 3) a single convolutional neural network (CNN) to be used for classification amongst all subjects. We have limited our preliminary results to EEG signal analysis of a left/right/rest motor imagery task as this is the most popular signal used in BCI applications. We have previously shown [3] that our proposed CNN can reliably decode intentions utilizing same architecture for multiple subjects. Here, we are extending our method to 4 new subjects and show that drastically reducing channels has insignificant effect on decoding results. We have also expanded our decoding results to 3-class classification and obtained the same decoding accuracies by only using a few channels for classification.
Our results show that drastically reducing the complexity of data, can still yield comparable performance while using a single decoder for multiple subjects. Since the choice of the channels for decoding is based on the mental task, one can envision the use of these methods to create practical, reliable online BCI solutions. Also using raw data for analysis and the use of a single architecture to classify all subjects allows for a hardware to be designed to even further improves efficiency of the system.
