Slashdot Mirror
The Baby Bootstrap?
An anonymous reader asks: "Slashdot recently covered a story that DARPA
would significantly cut CS research. When I was completing graduate
work in AI, the 'baby bootstrap' was considered the holy grail of military
applications. Simply put, the 'baby bootstrap' would empower a computing device to learn like a child with a very good memory. DARPA poured a small fortune into the research. No sensors, servos or video input - it only needed terminal I/O to be effective. Today the internet could provide a developmental database far beyond any testbed that we imagined, yet there has been no significant progress in over 30 years. MindPixels
and Cycorp seem typical of poorly funded efforts headed in the wrong direction, and all we hear from DARPA is autonomous robots. NIST seems more interested in industrial applications. Even Google
is remarkably void of anything about the 'baby bootstrap'. What went wrong? Has the military really given up on this concept, or has their research moved to other, more classified levels?"
I just got back from a workshop on this very subject, but nobody uses the term "baby bootstrap". It is now called "Developmental Robotics", and encompasses embodied agents, machine learning, and other biologically-inspired metaphors.
There is now a website dedicated to the idea. See http://DevelopmentalRobotics.org/ and http://cs.brynmawr.edu/DevRob05/ for a collection of papers on the subject.
Except computers never learned chess; humans programmed complex move analysis routines along with the rules, and many times a database of strategies with statistical weighting. There's a limited capacity to "learn: against opponents, but that's usually just more preprogrammed analysis and pattern matching than actualy spontaneous data linking. And like a poster higher up said, ther ewas a time we thought that was all one needed. It's not. We already have rudimentary AIs in labs that can "learn" in the sense they can create accurate spontaneous data links. The human brain (or the brain of any semi complex organism, really) is a black box with such unimaginable gears inside we're fumbling in the dark. It's hard to reverse engineer a mind becuase unlike reverse engineering a BIOS or widget, we don't really understand how a mind works, is put together, or even what it's really comprised of.
jX [ Make everything as simple as possible, but no simpler. - Einstein ]
The expert systems people hit a wall in the mid-1980s. An expert system is really just a way of storing manually-created rules. And those rules are written with great difficulty. There used to be expert systems people claiming that strong AI would come from rule-based systems. (Read Feigenbaum's "The Fifth Generation"). You don't hear that any more.
Hill-climbing systems (which include neural nets, genetic algorithms, artificial evolution, and simulated annealing) all work by trying to optimize some evaluation function. If the evaluation function is getting better, progress is being made. But what this really means is that the answer is encoded in the evaluation function. If the evaluation function is noisy (as in, "does the creature survive") and requires major simultaneous changes to make progress (as in "evolutionary jumps"), hill climbing doesn't work very well. There is progress, though. Koza's group at Stanford is moving forward, slowly.
The formal logic people never made much progress on real-world problems. Formalizing the problem is the hard part. Once the right formalism has been found, the manipulation required to solve it isn't that hard. There's not much work going on there any more.
The reactive robotics people also hit a wall. Literally, as every Roomba owner knows. Reactive control will get you up to the low end of insect-level AI, but then you're stuck.
Reverse-engineering brains still has promise, but we can't do it yet. Progress is coming from trying to reverse engineer simple animals like sea slugs. (Sea slugs have about 20,000 neurons. Big ones.) Efforts are underway to completely work out the wiring. Mammals are a long ways off.
Lately, there's been a trend towards "faking AI". This comes under such names as "social computing". The idea is to pick up cues and act intelligent when interacting with humans, even if there's no comprehension. This may have applications in the call center industry, but it's not intelligence.
I run one of the DARPA Grand Challenge teams, Team Overbot. On a problem like that, you can definitively fail, which means there's the potential for real progress. That's why it's worth doing.
>> By Bayesian spam filtering, I think you mean general classification problems, in which case, yes, neural networks can implement classification - it's a stretch to say that McClelland and Rumelhart's did, because the possible output included most non-repeating combinations of English phonemes and is thus nearly infinite, but the principle is there.
IIRC, mathematically it's been shown that neural nets and bayesian learning systems (such as spam filters) are entirely equivalent. Check out some of the work by David MacKay at the University of Cambridge.
The author of this post asserts his moral rights.