MIT researchers and their colleagues have developed a new computational model of the human brain's face-recognition mechanism that seems to capture aspects of human neurology that previous models have missed.
The researchers designed a machine-learning system that implemented their model, and they trained it to recognize particular faces by feeding it a battery of sample images. They found that the trained system included an intermediate processing step that represented a face's degree of rotation—say, 45 degrees from center—but not the direction—left or right.
This property wasn't built into the system; it emerged spontaneously from the training process. But it duplicates an experimentally observed feature of the primate face-processing mechanism. The researchers consider this an indication that their system and the brain are doing something similar.
(Score: 1) by g2 In The Desert on Monday December 05 2016, @04:34PM
Uh oh.
(Score: 0) by Anonymous Coward on Monday December 05 2016, @06:28PM
Parsimony pretty much dictates factoring out the angle and lighting on a face* and using basically a 3D model to actually compare with. If dealing with thousands of faces, then that's more compact than storing many angles and many lighting conditions for each face.
For example, if we consider 25 approximate possible lighting angles/types (including point source and hazy) and 25 approximate facial angles (including being tilted up and down for each horizontal angle), then 1000 faces would require storing 625,000 samplings (1000 x 25 x 25). That's a multiplicative mess. It's less info to first extract a 3D model, removing angle and lighting, and then compare to 1000 3D models (actual instances).
Any score-based AI-ish training system that either punishes or limits storage is probably going to come to a similar "conclusion", and more or less store and construct 3D models for primary comparison. (Neural nets typically pre-limit storage, but genetic algorithms can be more flexible that way.)
I suspect this will be found generally true of any AI of 3D objects. Factoring out projection-to-2D combinations as a preliminary step will result in less storage needed. Storing one 3D model is a lot less memory than storing all or enough possible 2D projections (photos) of such a model under enough possible lighting conditions.
* Skin tones and eye-color should still be used in identification, but perhaps treated as a different process path.
(Score: 1) by Scruffy Beard 2 on Monday December 05 2016, @07:12PM
I would say that was a $5 word, except the definition says otherwise;
I had to look it up [merriam-webster.com].
(Score: 0) by Anonymous Coward on Monday December 05 2016, @11:33PM
Mathematicians and linguists seem to use it often to mean fewest words or symbols. In this case, it means less memory or less combinations.
Or
Parsimony: what a Parson pays in alimony ;-)
(Score: 0) by Anonymous Coward on Monday December 05 2016, @07:40PM
Facebook has been paying for this research for a while. Look up the sample work they did -- I think Sylvester Stallone was the example face model.
The goal was to have FB be able to automatically detect and tag people in photos that were on an angle, poorly lit, or less than symetrical (scars/jowls, etc) that can make one person loon potentially like someone else on an angle.
Why they care is not something they discussed.
(Score: 0) by Anonymous Coward on Monday December 05 2016, @08:20PM
How significant is this finding? People are really good at spotting patterns, even when none exist. Is this purely coincidence with us seeing the single similarity and not seeing the countless potential similarities which didn't happen (like finding the number 666 within pi), or is this actually significant in some way?
(Score: 2) by ledow on Tuesday December 06 2016, @08:47AM
More importantly, what happens when you feed it an image that ISN'T a face?
This seems, to me, to be the part where any form of "AI" falls down. Unexpected data basically throws up random results, whereas "intelligent" systems tend not to. And the AI is rarely given the option "I don't know, I can't see anything" as a returnable response.
Whereas, even the youngest of children will just frown at you if you ask "which picture is Daddy" when neither of them are.
I'm sure there's an algorithm for "is there a face-like feature in this image" and one for "is this face similar to any other", but actually conjoining them isn't directly possible, and training for both circumstances is harder.
Even if there is - extrapolating doesn't produce "That's not a face, that's a butt" in preference to "No face detected".
The problem I have with this is that it's not AI, and realising there's an intermediate step that determines angle doesn't mean it's doing anything even vaguely similar to real intelligences or especially us. It would seem, in fact, a necessary element if you plug in lots of faces at lots of angles to immediately find a compensating factor that results in success more than normal because it reflects quite what "angle" to perceive the features at.
Feed it an image with an upside-down face. Humans are distinctly bad at recognising that - you can flip features like lips, eyebrows, nose, etc. in upside-down images and human rarely spot it until you rotate it the right way up. If it "detects" such a face, it's not recognising images the same way a human is.