12/30/2005

Eigenfaces

By using principal component analysis on a sufficiently large set of face pictures (5,000 to 1,000,000), you can break the pictures down into their simplest constituents (aka, their "principal components") which can then be recombined to create any given face within the set with high accuracy (often >95%). The principal components are considered "eigenvectors" of the original high-dimensional set of faces, or often just "eigenfaces" for short. As you can see, they often look pretty creepy.

This approach has proven fruitful in robotic face recognition, and it leads naturally to the question of whether the human brain is doing something similar. The fusiform face area (FFA) is an area of the temporal lobe that selectively processes face information - perhaps it is constantly doing some form of PCA on incoming face-like visual data, such that it determines the minimum number of neurons needed to represent all faces (the number of eigenfaces), and can then recombine these eigenvectors through coordinated firing to represent any particular face.

While it may seem like an outlandish hypothesis, remember that wavelet and Fourier-like transforms can be seen in the visual and the auditory systems. Also consider that such a "PCA module" in the brain might also be used more generally to distinguish between very similar objects: sure enough, the poorly-named "FFA" shows increased activation when bird experts look at birds, and when car experts look at cars. (It appears we are all face experts.)

Things get even more interesting when you look at autism, a disorder where kids show problems in the way they relate to people (to put it generally). Autistics show neural activation in response to faces in areas outside of the FFA, perhaps because their PCA system has become less localized to the FFA. Fitting with this interpretation, the regions activated in autistics differ from patient to patient. This "generalized PCA" theory would complement multiple theories of autism including those that state autistics are very good at systematizing ("extreme male brain theory"), and good at paying attention to detail but have difficulty integrating information ("underconnectivity theory"). Of course, this is pure conjecture and a "generalized PCA" system has not been shown to be at work in autism or aspergers. There are also clearly many other factors, such as differences in the long-range connectivity between brain regions and possible differences in the mirror neuron system.

4 Comments:

Anonymous Anonymous said...

where did you get that image of the eigenfaces??

1/18/2007 01:21:00 PM  
Blogger Chris Chatham said...

Good question - I'd have to hunt for it now. It was from an academic paper (I believe from people at MIT), but I don't remember anything else about it.

Why do you ask?

1/18/2007 02:24:00 PM  
Anonymous Anonymous said...

You write: "By using principal component analysis on a sufficiently large set of face pictures (5,000 to 1,000,000), you can break the pictures down into their simplest constituents (aka, their "principal components") which can then be recombined to create any given face within the set with high accuracy (often >95%)."
Caould you post a reference for this? I am currently working on a paper in which this would be an interesting point o make.
Thanks for your help in advance!
Cheers
Franzef

2/17/2007 02:07:00 AM  
Anonymous Anonymous said...

Saw some people here looking for information about PCA. I would highly recommend that you check out Lawrence Saul's fantastic presentation on non-linear dimensionality reduction from the Institute for Pure and Applied Mathematics 2005 Graduate Summer School. He gives the clearest explanation of PCA I've seen anywhere on the web, and then goes on to explain far better modern techniques developed since 2000.
http://www.oid.ucla.edu/Webcast/ipam/

5/21/2007 10:44:00 PM  

Post a Comment

<< Home