Cockroaches to cortex

Oh dear. Things aren’t going terribly well, are they?

            As we have seen, I can’t directly program Lucy’s computers to act intelligently, because minds are not like computer programs. In any case, behaviour that has been explicitly programmed in is more or less the exact opposite of intelligence – when people behave as if they are blindly following rules we don’t regard them as intelligent but call them zombies. What I need to do instead is to emulate the brain directly, using a neural network, so that intelligence arises as an emergent phenomenon that develops all by itself, through learning, rather than being hard-wired by me. But the conventional approaches to neural networks, using idealised neurons arranged in simple, usually non-recurrent ways, are hopelessly inadequate and have almost nothing in common with how real brains work. Evolution, it is true, can build messy, recurrent networks that have slightly more biological plausibility, but unless I’m willing to wait a while for them to evolve – say five hundred million years or so – I would be restricted to no more than a handful of neurons. It’s a good thing I knew about these problems in advance, or I might start to get depressed!

            Of course, I’ve been telling you all this bad news because I am trying to lead up to something. By building Lucy I’m taking a specific approach to AI that makes sense to me, after many years of thinking about these issues, but it may not make sense to you unless I point out some of the problems with the more traditional or obvious ways of doing things. I don’t want to bore you with too much background, and I know I promised you robots and other glamorous and exciting things, but the homework is nearly over. Before I heat up the soldering iron, I just want to tackle one more misconception that will help to explain exactly what I’m looking for.

Copyright © 2004 Cyberlife Research Ltd.
Last modified: 06/04/04