Re: Curriculum for AI

From: Cliff Stabbert (cps46@earthlink.net)
Date: Wed Jan 01 2003 - 11:05:47 MST


Ben Goertzel:
BG> I do not think Michael's tests are anywhere near adequate as a
BG> training regime for a baby AI.

Michael Roy Ames:
MRA> It depends on the design... but I get your point. Many AGI design types
MRA> will require a large quantity of experiential learning. I have not
MRA> attempted to describe that because different designs will require
MRA> different lessons and learning situations. Instead the Curriculum
MRA> focuses on what might be called 'book learning'. Lessons that every AGI
MRA> is going to have to learn in order to make progress. I will make this
MRA> clearer in my next version.

I think the basic outline of your curriculum is well thought out, and
I look forward to future, more complete versions.

However, I'm not certain what level of learning ability represented
will require training; I suspect some of the lower levels of learning
will be built into any AGI hierarchy. I do think the curriculum
offers an intriguing view of the successive levels of abstraction
involved in thinking. In particular, I can imagine an intelligence
having an "aha!" moment during lesson 14 and at various points during
the sublessons of 16.

I think the curriculum's ultimate utility may be as an AI methodology
testing ground: a test where "you can't accidentally hardcode the
relevant heuristics in" (as has happened in some past AI projects) --
and possibly, depending on how the curriculum develops, a test where
you can't do that even on purpose (because the tasks involved reach a
certain level of complexity). If the latter is possible it could
lead to a more easily automatable, and arguably more objective, test
than the current forms of the Turing test.

The lessons of Semester 1 seem the most useful and coherent to me, so
far. I'd like to see Lesson 16 fleshed out/expanded into multiple
lessons. I'd be very interested in lessons which explored algorithmic
design in more detail, e.g., provide sample datasets of input and
output, with the task being to create an algorithm; further lessons
that value the algorithm with the fewest steps, etc.

Semesters 2 - 4 might be more useful if the AGI is given some sort of
embodiment in the microdomain -- some way of effecting change and
observing its results. E.g., if it can build a mental model of the
"physics" of a microdomain by experimentation, in order to solve a
microdomain task.

You're doing a great job with this. If fleshed out and automated, I
think your curriculum could become an important tool for AGI
development.

--
Cliff


This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:41 MDT