Wow, that's improved amazingly in quality since the last time I saw something similar. Then, the 'reconstructed faces' were blurry blobs that were barely recognizable as faces. These are great.
It's an apples-to-oranges comparison though. These researchers generated 2000 faces. It sounds like from there they essentially did nearest neighbor search to find the predicted face. It's still impressive (and accomplished what they wanted), but it doesn't solve the problem of recreating the images from just the brainwaves.
From the article "In fact, they were nearly indistinguishable from the actual photos shown to the monkeys." It doesn't really sound like they did that.
They previously identified six "patches" in the monkey brain that are specialized for processing face-like stimuli. These are embedded in a much larger brain area involved in object recognition, but it's not clear whether faces are a special case or if other sorts of object patches remain to be discovered.
For invasive data like electrophysiology, not for fMRI. In fact few months ago there was a similar study about reading human faces from human brains via fMRI: https://twitter.com/ccnlab/status/866548346751725568
Somewhat related, this reminds me of a study with cats where they were reconstructing arbitrary black and white videos from visual cortex brainwaves at UC Berkeley https://youtu.be/piyY-UtyDZw.
I wonder if different social animals will have biases for certain facial features unique to their species, similar to what the end of the video suggests. It is what I would expect if our visual information system does work like a covnet, since important higher but level features would be pushed further down the stack, and inputs from our eyes that are similar would be "boosted". But I imagine the truth is much more complicated.
Here they don't seem to be experiencing any "monkey-like" feature boosting, so the area they chose is either low level enough, this is not a real effect, or monkeys are similar enough to human faces to not learn different low level features.
At the risk of bing flamed, while I read this and marvel at the technology, I wince at the terrible cost - animal testing - to achieve this.
CalTech's statement on animal testing is on their website [1] though I'm always skeptical of "oversight" groups established by an industry itself, as are PETA [2].
Could advances in this field be used to help blind people to "see" faces etc. by bypassing the eyes and instead directly communicating with the relevant neurons?
I would love to read the research paper but unfortunately it is Elseviered!!!
23 comments
[ 41.4 ms ] story [ 384 ms ] threadI wonder if they can do the same thing to a monkey that's sleeping and thus capture what faces a monkey sees in dreams?
They previously identified six "patches" in the monkey brain that are specialized for processing face-like stimuli. These are embedded in a much larger brain area involved in object recognition, but it's not clear whether faces are a special case or if other sorts of object patches remain to be discovered.
Looking at those pictures, I'd say that the monkey brain puts on 10 lbs.
And vice-versa.
I wonder if different social animals will have biases for certain facial features unique to their species, similar to what the end of the video suggests. It is what I would expect if our visual information system does work like a covnet, since important higher but level features would be pushed further down the stack, and inputs from our eyes that are similar would be "boosted". But I imagine the truth is much more complicated.
Here they don't seem to be experiencing any "monkey-like" feature boosting, so the area they chose is either low level enough, this is not a real effect, or monkeys are similar enough to human faces to not learn different low level features.
CalTech's statement on animal testing is on their website [1] though I'm always skeptical of "oversight" groups established by an industry itself, as are PETA [2].
[1]: https://iacuc.caltech.edu [2]: https://www.peta.org/blog/labs-gold-standard-seal-mean-anima...
I would love to read the research paper but unfortunately it is Elseviered!!!