Sunday, July 17, 2005

Face Cells

The June 23, 2005 issue of Nature reports work by Christof Koch and colleagues on the existence of "face recognition" cells in the hippocampus of the brain. This paper got a lot of play in the popular press because some of the cells only responded to famous people such as Halle Berry. The group found that the cells were highly selective to various views of a given person but not to another person. I think this work confirms some current theories of memory (see for example McClelland et al. Psychological Review, 102:419 (1995)). It's also more proof that there isn't much difference between humans and other mammals.

It is known that cells in the hippocampus in the rat code for spatial location in the same way. A given cell will only fire when a rat runs through a given spatial location in a given environment. When the environment changes, that same cell will then code for a completely different location. Location is important to a rat, just as the recognition of people is important to humans.

When the hippocampus is removed, humans can no longer form long term memories. They can remember things as long as they pay attention to it but once they lose their train of thought, the memory is completely gone. It is thus thought that the hippocampus is a form of mid-term memory that stores lots of information that is then slowly uploaded to the cortex for longer term storage.

It's useful to have different memory systems for different time scales because every time you remember something new you run the risk of erasing something old. One way out of this conundrum is to separate long term memory from short term memory. Simplistically, your hippocampus would store whatever information comes in and indiscriminately overwrite old information. Then slowly over time, the hippocampus would upload information to the temporal cortex (perhaps during dreams) which would update its synapses in a controlled fashion making sure not to erase important old memories.

What this paper shows is completely consistent with this idea. From theoretical work on associative memory, we know that the capacity of any neural network is limited by how correlated the stored patterns are with each other. The more correlated the patterns, the more likely they are to interfere. Thus, one way to make sure you don't overwrite old memories is to make sure the input patterns are orthogonal. The hippocampus may serve this purpose. A very sparse code, where only a few neurons encode a given concept (like Halle Berry), automatically orthogonalizes the patterns representing given memories presented to the higher cortical areas.

A sparse code is not robust because if you knock out that particular neuron you lose the memory it coded. A more robust code would be a population code where a large group of neurons encodes a given concept. The problem with this type of memory is that it's hard to train a network. So the way to overcome the trade-off between robustness and speed is to have a fast but fragile system (hippocampus) feed inputs to a slow but robust system (temporal cortex).

It is known that inferotemporal (IT) cortex of monkeys also respond to faces among many other percepts and that a given cell in IT will respond to a wide variety of images. So, if they ever get a chance to implant electrodes in the temporal cortex of humans, I'm sure they'll find similarly behaving cells.

3 comments:

Anonymous said...

alright carson, i thought you would jump on this one and i'm afraid i disagree.
indeed, this study (made on eight epileptic people) seems to be a strong argument for the one-cell sparse code (formerly known as grand-mother cell). I have nothing against sparse code as it is but i'm strongly skeptical of the equation 'concept=neuron' the way you mention it.
First, I can find one million reason that one neuron answers to a particular type of stimulus (among the few that were presented) and not to the others without giving a label to that neuron.
Second, this is the same problem with place cells - we do observe strong correlation between position and a particular neuron ; fine. but to assume that this neurons 'codes' for the location is to fall in anthropomorphism (french direct translation i do not know it that exists ...) trap, i think.
Third, let's assume you've got an Halle Berry neuron where does it project ? does everybody have one ? so perhaps have we one neuron for each people we saw the face of ?
all these questions we cannot even answer for a rat's place cell (notably because of the very slow firing rate involved)
i did not like this paper because it did not answer any good questions and put arguments in the bag ot the good'old AI symbolist camp.
./ed (aka hedi)

Carson C. Chow said...

Hedi,

I didn't mean to say it was a one neuron code but a sparse code, say a few percent of the neurons coding for a concept and in the hippocampus only. The data in animals shows that this is probably the case as well for independent maps. However, the higher you go in cortex the larger the receptive field implying more of a population code.

No, everyone does not have a Halle Berry neuron. What it implies is whatever you happen to be exposed to at that particular moment is stored in hippocampus. So things like what you are reading, who you see in the news etc. This is stored in a sparse code in hippocampus. Then over time what is important gets consolidated into long term memory. I think this paper merely confirms what we already know to be true in animals. I actually wrote the post as a mild critique of the hype around it. Unfortunately, it seems as if I failed completely.

cc

Anonymous said...

carson,
sounds convincing
i wasn't much aware of that hype around the paper from here
./ed