I have had this idea that i first wrote about in this arxiv paper:
Its to use a generative algorithm to produce fake fMri data conditioned on the environment to controll a humanoid robot.
I have come some way since then with the development of a way for representing the fmri data in a compressed manner and more importantly for having a consistent representation that cuts accross many brain wiring types.
We could simply have a dataset of humans learning something while taking brain scans of their brains,and having them do excersizes conditioned on questions, then generate these type of fMRI images in a generative algorithm and train it by conditioning its evolution on learning material before asking it questions.
What do you think?
Anyone want to work on this with me?