Increasing the efficiency of machine learning methods utilized in brain computer interface systems
01/30/2020
András Adolf1, 2, Gergely Márton1, 3, István Ulbert1, 2
1 Institute of Cognitive Neuroscience and Psychology, Research Centre for Natural Sciences, Budapest, Hungary
2 Faculty of Information Technology and Bionics, Pázmány Péter Catholic University, Budapest, Hungary
3 Doctoral School on Materials Sciences and Technologies, Óbuda University, Budapest, Hungary
The vast majority of non-invasive Brain-Computer Interfaces (BCI) works with electroencephalographic (EEG) signals. After processing this data, these interfaces are capable of issuing command instructions that can be used to handle an arbitrary device. Since the recording of EEG data is cumbersome and basically these patterns can vary greatly from person to person, there is relatively little data available. When we want to classify with a machine learning system, there is a need for a great dataset. This is not provided in many cases, so in our work we primarily focused on this problem: how to teach effectively classification systems with minimal amount of input data. In this work first we have created and taught a relatively well-performing neural network on the Schalk EEG database. The structure was a 3-layer convolutional network, followed by 2 fully connected classification layers. The network receives a 3D array of EEG signals as an input: the data is organized into a structure corresponding to the two-dimensional spatial arrangement, the 3rd dimension is time. In the case of 5 classes, the ratio of successful classifications was around 50%, while in the case of a simplified task for the classification of 2 classes this ratio was around 80%.We have implemented methods capable of learning from minimal amount of input data. To simulate this, the size of the dataset was reduced, the system performed around 45% of accuracy. That was improved to 55% via transfer learning and 54% via data augmentation, which is a remarkable development.