Brain-machine interface spelling device based on reinforcement learning

The current demonstration will show a novel EEG-based brain-machine interface (BMI) spelling device. It combines real-time decoding of brain activity signals with a reinforcement learning approach to rapidly infer the characters the user wants to write We have developed a communication interface based on multimodal signals that allows users to communicate using different input devices depending on their condition. The proposed solution is an enhanced version of classical matrix-based systems in which machine learning techniques are used to speed up communication and reduce the user’s workload. Importantly, the implementation of these techniques also takes into account the speed at which the user can deliver the input and ensures that user’s input mistakes have small impact on the communication performance. The interface is composed of a character matrix, in which a moving cursor automatically scans the characters. At the same time, the user gives feedback to the device about the correctness of the movements. Contrasting to conventional systems, the cursor does not move in a pre-defined manner; instead it moves towards the most probable character to be written. This probability is estimated based on a language model and the feedback provided by the user.