Skip to main content
  • Expression of Interest

    Non-hardware specific artificial intelligence for powering noninvasive brain-computer interfaces
    neural networks, deep learning, brain-computer interface, artificial intelligence

    Today's EEG signal processing algorithms are hardware-specific. State-of-the-art neural networks trained on a specific dataset obtained by a single type of EEG device are incompatible with inputs generated by other types of EEG devices. Whereas the AI agents used in BCI systems for signal processing should be more robust: they should be able to process signals from various devices, similarly to how the neural nets utilized for image processing are robust to the variance of the types of cameras and lighting conditions. The main benefit of the herein proposed project is that the companies and research institutes within the participating countries will have the most advanced tools (in the form of neural networks) for the classification of mental commands and detection of mental states and responses from EEG signals. Frameworks will be established for data collection, storage, neural network training and testing. These frameworks can remain in effect after the funded period so that future advancements in AI research can be continuously incorporated. The neural nets can be applied in diverse product lines such as BCI headsets which include on-device inference, and ones which are streaming the recorded signals for real-time processing into mobile devices, desktop and laptop computers, and into the cloud. Our first intermediate goal is to create a large unified Brain-computer Interface (BCI)-related EEG signal database, which will allow us to train the most modern neural network types with superior efficiency. To achieve this, it is important that the networks should be able to process data obtained by different recording hardware types. The input of the artificial neural networks will be supplied through initial adapter layers, in such a way that each input channel of the main neural network will be specific to a location on the scalp. This construction will render the networks suitable for receiving data from the different headsets, via device-specific remapping and interpolation layers. Supervised and unsupervised machine learning methods will be applied using the large volume database of EEG signals yielded by standardized measurements.

    {Empty}
    {Empty}
    {Empty}
    {Empty}
    {Empty}
Search