Meta-BCI: Perspectives on a role of self-supervised learning in meta brain computer interface

A time-efficient process for building a decoder transforming neural signals into intended commands must be considered equally significantly as maximizing its interpretation performance, in order to establish a successful brain-computer interface (BCI) system. Many decision neuroscience studies revealed evidences for, at least in a task of reinforcement learning, the theory of human decision process consisting of two latent strategies: model-based (MB) and model-free (MF) – goal-directed and habitual manner, respectively. Several BCI studies utilized such findings to present how the overall system can be advanced by employing the pre-built fMRI-based neural computational model as a supervisor for constructing an EEG-based latent strategy decoder. However, such a method demands building a supervising model in advance, which is both heavily time-consuming and computationally costly. Additionally, human behavioral experiments may be additionally required for each different task – highly uneconomical. In order to reasonably deliver our perspectives on a solution for the given issues, we first clarify a definition of meta-BCI – any BCI system that is composed of a latent neural strategy decoder along with an intention interpreter. Subsequently, a few distinct attributes of existing meta-BCI systems in comparison to the conventional ones are described. Last but not least, we suggest how the meta-BCI system can be established both computationally and time efficiently by building a latent strategy decoder in a self-supervised manner, along with our future study plan on the meta-BCI in the particular context. Arguably, a well-established meta-BCI would not only allow swift and precise translation of neural signals into, for instance, movement kinematics with low computational cost, but also offer a great potential for efficient and accurate interpretation of various high-level cognitive intentions from neural representations.