That's wild! I wonder why they need to do the bit where they merge clips from youtube though. I would have thought that since you can predict which images match the neural activity best, you could build a "most likely scenario" image without mucking around with the youtube reference images. (although these people are way smarter than I, so I'm sure there's a very good reason)

Oh, ok, youtube videos. Was wondering why there were suddenly buildings in the background of some of those.

My guess is (since brain images work with spatial ideas instead of pixels) that comparing the brainscan's "idea" with the youtube video's "idea" to produce an image that has very little to do with the pixel content. fMRI's have a very low spatial resolution, and an even worse temporal resolution (sample rate) because they measure bloodflow in the brain, as opposed to actual neural activity (hence the once a second sample rate).

I suppose the emphasis of this particular achievement was that they could get a plausible "video-feed" from fMRI data.

I wonder what kind of quality could be acquired from implanted electrodes.

There was already something like this before except they displayed actual pixel output and it was low res black and white (IIRC). Seems like they may have improved the same technique by being able to get colour as well (possibly not, may still be black and white) and are probably just using the video matching to improve the image even more (instead of displaying pixellated noisy output compare to many videos and pick the top 4 matches and find that it roughly matches the actual video shown and looks even better).


Here's something else you might find interesting (basically the reverse- sending pixellated video of the world to the brain for a blind man): http://www.youtube.com/watch?v=TW78wbN-WuU&t=85 (the rest is interesting too but it's linked to the specific part of the video)

We do need to keep in mind that this doesn't really mean that we can retrieve an image from somebody's brain, but instead that we can make a vague guess at what the image is assuming we know all of the possibilities. That isn't to say that this isn't really cool research, or that the technology won't develop beyond this, but it's not quite as sophisticated as it seems. As zalo said, though, I wonder how much more accurate this would be with direct neural implants, since the main issue seems to be how slowly the fMRI scan (or, more accurately, the blood flow in the brain) updates.