Believe it or not, but I am actually teaching kindergarten kids right now. While some might consider this a big step down from the university teaching I was doing two years ago, the experiences I’m confronted with on a daily basis really keep my mind churning.
In addition to cleaning up plate loads of snot that oozes from a plethora of orifices I also have the time honored ritual of teaching children abstract concepts like colors and how much wood would a woodchuck chuck.
As a result, I’ve begun to rethink a bit of how strong and weak AI approaches have developed over the years. I mean how do you teach a machine something — how do you get it to learn? If you were an engineer starting from scratch, how would you approach this method?
After all, humans are squishy, brittle machines, yet it appears that it only takes a relatively short period of time to condition necessary parts of the brain in order to detect patterns and rules for classifying information, like the color red.
Exhibit A
I am assuming my sample size is big enough: a couple dozen middle-class Taiwanese kids between the ages of 3-6. Over the last few days several new kids were enrolled in my class. None of them spoke English. Within a day I got one of them to recognize shapes like circles, squares and triangles. And now after a week, not only have do they all know my name and the names of others, but also the weather (hot, cold, snowy, rain), basic colors, types of food and a number of other relatively simple concepts.
Sure they could be savants, but based on the “normal” distribution curve and observations from other teachers, all of the other kids at the schools (I teach at 3 different ones) progressed at about the same rate. In fact, based on the academic research into stages of cognitive development, they are smack dab on target with the rest of their human peers.
This is in stark contrast to simply attaching a webcam into a workstation and expecting the machine to somehow magically distinguish your face from mine. The hardware for this particular application arguably can discern curves, lines, colors and distortions but the software end, including the device drivers, has to be developed with the necessary code to translate the totality, the whole sum of the parts.
Ultimately you and I are also auto-didacts, capable of teaching ourselves new concepts, a feat that is still beyond the reach of most AI applications.
Thus, I suspect that a balance between hardware (akin to Minsky’s Thinking Machine) and software is where cognizance, sentience and self-learning resides. Thus, you need a machine(s) fast enough to be able process numerous inputs in real-time, yet software robust enough to tackle fuzzy, seemingly abstract variables (e.g., even something relatively benign like where does red stop and infrared begin).
IBM’s Blue Brain project appears to be making quick progress on the hardware “strong” side of this equation.
While it is a somber read, Wired magazine recently published an interesting piece on the lives of two AI researchers who died last year. With any luck, their novel “weaker” approaches to teaching machines and getting machines to think will live long into the future.
In addition you may be interested in the innovative work conducted by Luis von Auhn. I’ve mentioned him before as he is the brains behind the “Captcha” as well as Google’s Image Labeler. His Google Talk on Human Computation is also outstanding.
See also: Machine learning
Natural language processing
Speech recognition
Computer vision
Mechanical Turk