townx - Neats vs. scruffies - Comments

There's definitely space for

elliot — Mon, 16 Jun 2008 15:29:50 -0500

There's definitely space for both.

I've been playing with RRobots this week, which is about coding a software robot to a low-level API. You build higher-level rules on top of fairly low-level ones (turn gun, turn radar, accelerate, turn body, fire).

What's interesting to me in that case is the range of solutions: I found one on the web which basically used trigonometry to work out where to move and fire at the opposing robot. In my opinion, that's not interesting in this domain: my preference was to not compute, but to swing the radar around vaguely, move semi-randomly, use a more heuristic approach. For me, that's the right level of description for that kind of fight or flight behaviour.

But for language processing, I want the system to provide an interpretation. I think abstracting down into the low levels (ironically) perhaps makes it more difficult to work back up to a high-level description (e.g. in human language). Would Google work so well if they'd programmed it using neural networks? It's a tough question, that's for sure.

Scruffy

Natalie Andrew — Sat, 24 May 2008 08:50:20 -0500

I agree with your points about neural networks, they're messy and look just as indecipherable as the things they're trying to model. On the other hand, I like the idea that the 'how' is embedded in the nature of the network. The simple nebulous action of adjusting weights gives rise to a concrete rule. I don't know, is reasoning really that defined in humans? How much is illogical? If it's a practical machine you're building though I guess you have to know how it got to the answer.

I put myself in the scruffy camp too. That's mostly through laziness. I find that I can implement complex chemical and cellular behaviours by following the descriptions of what they might be doing rather than having to write and solve an ODE. At the very least it helps me to understand how I'm getting a new behaviour by tweaking the rules. Unfortunately I think that won't cut it if I try to publish lol.

Neats vs. scruffies

elliot — Thu, 20 Mar 2008 15:39:57 -0500

I did my Ph.D. in artificial intelligence, so was interested to read a few Wikipedia articles about it. One distinction I'd never heard of was neats vs. scruffies in the field.

I put myself in the scruffies camp, probably, though I always had a yen for predicate logic and formal grammars. To my mind, some of the AI scruffies weren't scruffy enough, and tried to model human intelligence without any reference to psychological data. I tried to redress the balance a bit, and compared my program's output with psychological data on human inference during story comprehension. You can read all about it here.

At the time I did my Ph.D., I was pretty unfashionable, as I was researching symbolic AI approaches, while everyone around me seemed to be doing neural networks. However, I thought that while sub-symbolic approaches might produce intelligent output, I struggled to see how that would lead to a description of the solution, or anything that might be built on or added to by humans. If you're trying to program a reasoning system, for example, is it enough to train a neural network to create associations, or do you need to write something which can reflect on the process by which it reached its solutions? Neural nets are great for recognition tasks, but I was never convinced they were suitable for reflecting on how they completed the task. I'm sure there are plenty of counter-arguments to my limited opinion, so feel free to enlighten me.