notes on combining logical domain knowledge with statistical reasoning, focusing on EBL

* explanation-based learning (EBL): popular in the 80s, nowadays very out of fashion (like most logic-based AI)
* the idea, I think, is that logical theories about a domain are refined based on data. The idea is maximize bias, i.e. use domain knowledge to create a strong prior for learning.
* related to ILP, except that the latter seems to focus more on concept discovery (though I really should look this up on AIMA)
* schema acquisition is similar to production compilation. (There are 2 separate connections between EBL and DOP: (1) Rens Bod's idea of modeling chunking of scientific derivations (2) Khalil Sima'an's PhD thesis)
* probabilistic EBL: published in 2006. Does it have potential to make EBL hot again?

People, papers and systems:
William W. Cohen (1992): Compiling prior knowledge into an explicit bias
Gerald DeJong (I like the two paragraphs on "What is EBL?". He has recent publications on EBL!)
Pat Langley - Unifying Themes in Empirical and Explanation-Based Learning
Paul Thagard's Bacon
Mike Pazzani and the Occam system

I love Dave Marshall's notes (on EBL), especially Learning II. His notes are high-level comparisons between different approaches. Google keeps pointing me to them.

Eliezer Yudkowsky: illusion of transparency

Earlier this week, I followed several links from this post by Eliezer Yudkowsky to previous posts of his. I've always known that he was a very smart guy, but that day, I really came to admire his mind. Here are some of the posts:

Expecting Short Inferential Distances: people tend to oversimplify things because they are skeptical of complex explanations.

Making Beliefs Pay Rent is about theoretical terms. It reminded me of why I'm not a mathematician.

Guessing the Teacher's Password

and my favorite:
Illusion of Transparency: Why No One Understands You, which is about theory-of-mind, and how we are all deficient at it, since we assume that others see what we see (because that's the easy thing to do cognitively). At least one of the Keysar papers he linked to (out of 3) was fascinating. I can't tell which one, since my Acrobat is broken.

I left a comment:
<< Eliezer, the so-called "expert blind spot" is IMHO one of the most important problems in lecture-based education and even scientific communication. One of my dreams is to make AI that helps address this, via cognitive models of both experts and novices. This is one of my holy grails.
The AI would understand what was said, and "translate" the message to each novice individually, taking advantage of their pre-existing knowledge. In some cases, this "translation" would involve lengthy tutoring with new concepts and knowledge. >>