Object-responsive, abstract, neurons in medial-temporal lobe
A fascinating study finds neurons which respond selectively to certain individuals or landmarks in varied representations. For example, one unit responded to “various pictures of the actress Halle Berry - as well as drawings of her and her name written down”. Another one responded both to pictures of the Sydney Opera House and also to the letter string “Sydney Opera”.
For most of the responsive units, the experimenters could only find a single concept which the unit responded to. The authors don’t feel that the units are “grandmother cells”, but rather, they think they just didn’t test for the other things that the cells would have responded to. Still, this argues strongly for pretty sparse coding.
R. Quian Quiroga, L. Reddy, G. Kreiman, C. Koch, I. Fried. Invariant visual representation by single neurons in the human brain. Nature 435, 1102-1107 (23 Jun 2005) Letters to Editor.
July 20th, 2005 at 12:14 pm
Ah, another grandmother cell article. (OK, Bayle was good about specifically saying it’s not a GC article, but almost all the popular press seems to think otherwise…) I think the grandmother cell idea has come full circle it’s become so “out of fashion” that now it’s back in fashion!
Or has it?
The original reason that the grandmother cell idea was viewed as preposterous was that even though we seem to have a lot of neurons in the brain — O(billions) — we can recognize a seemingly infinite set of objects. But, perhaps you might argue that a few billion different objects looks essentially like infinity to us. Okay, but there is still a further more troubling implication of the GC theory: If an object is represented by just one cell (as sparse as one can imagine without resorting to synapses), then we might worry that the death of a single cell can totally wipe out the knowledge about a particular object. Does this every really happen? My intuition is no but I’d love to know if there’s any applicable research to think about here.
Anyways, I should say that I think Leila and Christof are great scientists and the study itself is without a doubt very, very interesting. It’s nice to see that something like a GC theory but less extreme (ie. sparse coding) can explain this kind of high-level visual recognition. To me, it’s intriguing that the representation at this level wouldn’t be ultra-distributed.