Basically, they had this experiment where they showed the subjects 400 12x12 (I think) pixel images of letters and measured various things in their brain.
They used (most likely) neural nets and machine learning to draw conclusions about what was going on, and then they showed the subjects a series of NEW images that were not in the training set, measured the brain activity, and they were able to display a graphic that had white, grey and black pixels that closely resembled the white on black lettering that they had trained it on.
They were not exact pictures, but kind of like fuzzy images that you could see the whiter areas were were the letters were, and the darker areas in the images were aslo where the background was.
Interesting stuff, but I think a far cry from the reading people's dreams they believe is possible in the future.
They've established the basic principle. All it needs now is refinement.
This "basic principle" is no different when they do other MRIs that can detect patterns in brain activity (specific to one individual). The computers aren't doing any interpretation of brain activity here, just recognizing what it has been preprogrammed to to recognize.
What you're proposing is that we build a recognition database of every single brain pattern and its correlating image. Damn, what kind of computer does that remind you of? Oh yea - the human brain!
Not only that, but each individual has unique brain patterns, so this "computer" would have to be able to interpret a wide variety of signals. We're so far off from understanding the brain that this isn't happening, ever.
You lead the topic to a very big question. If you think here computer is not doing any interpretation, what you actually say is computer learning methods don't mean computer can think. In this article, scientists do make computer do some "interpretation" because the training data and test data are different. (NEW images and reconstruct).
People here have much more to do to fulfill the goal of "reconstructing high resolution, color image", but the problem is not computers do think or not.
Basically, they had this experiment where they showed the subjects 400 12x12 (I think) pixel images of letters and measured various things in their brain.
They used (most likely) neural nets and machine learning to draw conclusions about what was going on, and then they showed the subjects a series of NEW images that were not in the training set, measured the brain activity, and they were able to display a graphic that had white, grey and black pixels that closely resembled the white on black lettering that they had trained it on.
They were not exact pictures, but kind of like fuzzy images that you could see the whiter areas were were the letters were, and the darker areas in the images were aslo where the background was.
Interesting stuff, but I think a far cry from the reading people's dreams they believe is possible in the future.