RE: Parallelizing attention and credit-assignment

From: Ben Goertzel (
Date: Fri Sep 09 2005 - 15:02:07 MDT

> > IMO assigment of credit is a good example of an AI problem that no
> > one has figured out how to parallelize effectively yet.
> > Traditional AI
> > approaches to assignment of credit such as Q-learning, Holland's
> > classifiers, or Baum's Hayek are elegantly parallel in nature, yet
> > highly ineffective.
> Can you give some reason why you say this?
> This is important.

That these methods are highly ineffective is well-known. None of them
really works, which is why none has been used as the foundation of any
reasonably functional narrow AI system let alone AGI system.

The hypothesis that centralized techniques can work better is the
controversial one, and I haven't yet proved it.

The basic concept is that assignment of credit should be treated as a
data-mining problem. Keep a centralized record of all activities within a
system, including goal achievement. Then mine this record to learn patterns
of activity that are predictive of goal achievement, and do inference on the
mined patterns to guess better patterns than the data-mining algorithm is
able to find. The details regard what sort of data mining and reasoning
algorithms to use, and what sort of data structure to use for the
centralized record; the Novamente design contains specifics in this regard.

Basically, parallel, self-organizing assignment of credit algorithms are
trying to do this kind of data mining, but they're doing a bad job even
compared to traditional data mining algorithms, because they are trying to
do it via funky massively parallel self-organizing methods that don't work
and IMO make the problem seem harder than it really is.

The brain may well use some funky massively parallel self-organizing method,
but it has a lot of neurons to waste ;-)

-- Ben G

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:52 MDT