30 May 2007

Google: Not Big Brother (Yet)

Ars Technica reports that Google has surreptitiously added facial recognition technology to its image search; the function is apparently only available to those who know the secret search parameters. Currently the technology is useful just for finding pictures that have faces in them, not for identifying the faces, though the AT article predicts that it's only a matter of time. Facial recognition technology, of course, is one of those things that sounds really alarming from a privacy standpoint: If citizens are viewed by surveillance cameras 300+ times a day (as the average Brit is), and if these cameras were enabled with true facial recognition technology, well, any libertarian worth her salt can see how this works out. However, facial recognition technology is alot like karaoke - ridiculously easy to do, but very hard to do well. The concept is easy: give the computer a template for facial patterns, and show it lots of pictures until it understands how to find the face. Then, teach it to distinguish different faces using the same procedure. Heck, babies can do this almost from birth, even with other species (interestingly, adults get worse at it as they get older, and are much worse at recognizing faces in other species). But although there are many facial recognition programs around, they don't work very well to begin with, and are notoriously easy to fool with a different haircut, a hat, or a pair of Groucho glasses.

On the other hand, Google has a knack for solving difficult problems using surprisingly simple solutions. G. revolutionized websearch by eschewing complicated natural language processing options designed to understand exactly what the searcher wants, and instead found a fast, easy way to make a reasonable guess. And facial recognition is ultimately a pretty similar sort of problem -- like semantic language processing, it's accomplished by humans using a complex neural network, which is hard to replicate in a computer (so far). Teaching a program to simply make a decent guess might be a much easier, more elegant solution, which is Google's specialty.


Barry said...

Excellent point; Google might launch this on a limited basis using a 'guess' feature. Present a small set of hits, and let the human choose one. Repeat a zillion times, and a neural network might be very well trained. This would also provide lots of human imput to test numerous alternate algorithms.