Convolutional architectures for transfer learning in brain-computer interface applications

Abstract

Brain-Machine Interfaces (BCIs) may help individuals with movement restrictions to control devices by processing electrical signals captured by electrodes using, for example, commands to move a cursor up/down or right/left. The use of neural networks, especially convolutional networks, has shown promising results when used with the time series of electrodes as input. One problem is that training these classification algorithms requires a large number of labeled examples. One way to increase the amount of data available for training is to train models with data from multiple individuals. We can then use the model to classify signals from new individuals with the aid of transfer learning techniques. In this work, we will use convolutional networks to perform decoding in BCI tasks. To allow the use of data from multiple individuals, we will evaluate neural network architectures that allow sharing most of the parameters between subjects. A small part of the weights would be adjusted for each subject in order to capture individual differences. These architectures will then be used as a base for transfer learning, where only a small dataset from new individuals will be available for individual adjustments. We will compare the effectiveness of the proposed architectures with other state-of-the-art architectures.(AU)