Computationally Empowered Learning Strategies for Non-Invasive Intracranial Brain-Computer Interfaces

A brain-computer interface (BCI) is a system that is able to acquire brain signals and translate them into an input for a control system. Typical applications of BCIs span several domains, including the rehabilitation of patients with severe motor disabilities, prosthetic control, and augmentation of virtual reality environments. Most widespread BCIs are based on data recorded with an electroencephalograph (EEG), a device where a set of electrodes is placed on a subject's scalp to record the electric potential. Scalp EEG is a non-invasive and relatively cheap solution that provides measurements of brain activity with high temporal resolution. On the other hand, EEG is known to suffer from poor spatial resolution. This makes it far from trivial to reconstruct actual electrical activity inside the head volume, a technique known as EEG source imaging (ESI). For this purpose, a wide range of techniques have been developed in order to faithfully reconstruct active sources in the brain from EEG data. Nevertheless, it is still not quite clear whether these imaging techniques can prove to be generally useful in the context of brain-computer interfaces. Moreover experimental data is characterized from the lack of a ground truth on the source activity under imaging. This, in addition to scarce availability of standardized BCI datasets, leads to difficulties in prototyping, testing and comparing performance of ESI based BCIs (and of BCIs more in general). The goal of the present work is to investigate the usage of ESI techniques for the augmentation of brain-computer interfaces. In order to do so we have investigated different BCI paradigms and looked for the one where the usage of such techniques can improve the performances and stability of the system. In particular, we have engineered several state-of-the-art BCI pipelines and studied the characteristic of the biosignals related to the BCI paradigm of interest, working on different publicly available datasets in order to design a proper imaging technique. To ensure a scientifically sound and reproducible protocol, we have also carried out the recording of our own experimental BCI dataset, which validated our results and helped us envision new research directions. Faced with the problem of the scarcity of available data and lack of ground truth in EEG recordings, an often reported serious obstacle to progress in the field of BCIs, we have also investigated the application of generative modeling with deep learning strategies. In this context, we have developed models that are able to generate EEG signals with high fidelity. The generated data can then be used to augment and improve existing applications in a variety of downstream tasks.