Jerry Fodor; Proof-Theoretic Semantics

[gusl]

I like Jerry Fodor!

Would he be a pioneer of the so-called "computational turn"?

... a few link hops away, two Wikipedia articles: Proof-Theoretic Semantics (where the meaning is the use), and Logical Harmony

Btw, like Fodor, I am not interested in connectionist models... While I accept that intelligence must ultimately be implemented in terms of neural networks, I think it's more useful to think about phenomena at a higher level of abstraction, where logic plays a role. It's also more fun. It should also come as no surprise that I'm a fan of ACT-R.

What have we ever gained from NN modeling?

Comments

[fub.livejournal.com]

Unfortunately, there is no real way to know what your connectionist model is learning. Horror-stories abound.

When I was busy with my graduation project (I studied Cognitive Science in Nijmegen), my professor always told me that there were three ways to make intelligent systems:
- If you know how humans reason about the domain, then just implement that in software, and you're done.
- If you have a good heuristic, just implement that in your program -- it will work most of the time.
- If you have no clue whatsoever how to tackle the problem, just use a connectionist model! Train it a bit, and you'll always get some output...

Also, with the right knowledge representation, declarative systems might 'learn' as well, if you allow the knowledge base to be updated.

[bondage-and-tea.livejournal.com]

Nice hierarchy of understanding :-)

Not sure I agree with the "if you have no clue whatsoever how to tackle the problem, just use a connectionist model" -- don't you have to have a vague idea of how to represent the problem in order to get a decent net, for instance what features are salient for the problem.

I'd like to learn more about symbolic learning algorithms. I know of the likes of ILP (http://www-ai.ijs.si/SasoDzeroski/ILPBook/) and ADATE (http://www-ia.hiof.no/~rolando/), but have just played at the software level, not tried to understand the algorithms.

Also interesting is the hybrid connectionist-symbolic stuff, e.g. Pinkas's translation of his penalty of logic into hopfield networks. There's some interesting more recent stuff of this nature going on.