There is intense interest regarding object recognition in the cortex - it is a local or hierarchical distributed function?
This paper presents a method of decoding FMRI data into the original image viewed. IMHO - The results are stunningly good.
It was necessary to use hierarchy to recover the original image lending weight to the case for a distributed hierarchical representation.
Here’s what happens when micro-electrode arrays are used in a monkey brain to distinguish the response of single neurons to visual stimuli. A neuron fires orders of magnitude more quickly in response to synthetic, evolved patterns than to natural video. One neuron seems to recognize another monkey that wears a red tag; another seems to recognize a human caretaker who wears a blue surgical mask.
These results, too, are fascinating, to me at least.
Related:
Another group has produced similar manipulation of V4 neurons, proposing that this could become a tool for brain-state manipulation at the level of the individual neuron.
Such a proposal becomes even more interesting in light of research showing that “Emotion schemas are embedded in the human visual system.”
Exciting times!
This almost seems too good to be true!
Indeed! How could they produce such reconstruction with only a limited EEG electrodes? I suspect overfitting as an explanation.
From the article:
Researchers from Russian corporation Neurobotics and the Moscow Institute of Physics and Technology have found a way to visualize a person’s brain activity as actual images mimicking what they observe in real time. This will enable new post-stroke rehabilitation devices controlled by brain signals.
Since the device needs to be trained on each subject, the “post-stroke rehabilitation” will only work on people having already undergone the training before their stroke, right?
Yes, they would need to train models specifically for each person, because the patterns in each person will be different.
I’m running experiments of this style with NuPIC for my research – if anyone has use cases 
We tried to reproduce the paper Bitking linked in our brain hackathon this weekend
but unfortunately we didn’t get there. It remains to be seen if we keep the investigation up now after the hackathon.
I think the criticism here underlines the correct figure of merit: the approach can’t be evaluated simply by how well the generated images approach the original images, but rather how covariant they are. A causal measure of how much info is taken from the EEG.