From: Ben Goertzel (ben@goertzel.org)
Date: Sun Oct 24 2004 - 08:05:16 MDT
> A lot of the systems had impressive behavior, but most were
> dead end approaches, in my opinion, because they made logical
> reasoning fundamental with learning as an add-on. The most
> impressive talk from the main stream AI community was by Deb
> Roy, who is achieving interesting vision-language coordination
> with systems that are fundamentally about learning.
>
> It was good to see Ben again, and to meet Moshe Looks and Pei
> Wang.
>
> Cheers,
> Bill
Hey Bill --
Just a brief comment about Deb Roy's work., then some more comments on
human-level AI in general.
Deb's work is really interesting, however it actually represents a move AWAY
from learning, as compared to his thesis work a few years ago. Then he was
focusing on having his software system learn visuomotor groundings for
linguistic terms (nouns like "apple", etc.). Now he has been making his
system do more complex stuff, but via hard-coding control schema into it,
rather than via learning. When Moshe and I talked to him after his talk,
however, he said his next step would be to implement some approach to
learning these control schemata. But he didn't seem to have such a clear
idea about how he'd do it. When Moshe asked him about how he's implement
grounding of prepositional and subject-argument relationships (arguably the
nontrivial part of language-grounding), he said he didn't have an approach
to that yet because he didn't know any good way to represent that kind of
knowledge on the cognitive level.
So, I think his work has tremendous promise; but yet, I couldn't help wish
he'd pushed it in a more learning-based direction during the last few years.
On the other hand, I can sympathize, because -- for instance -- over the
last year we've had a Novamente team member (Mike Ross) create a hard-coded
language processing module. Why? Because we needed it for a commercial
project. (Deb Roy is an academic -- academics don't have revenue pressures,
but they often have demo pressures associated with funding sources!). Now,
during 2005 Mike will replace this hard-coded language processing module
with a learning-oriented language processing module. Basically, the need
for incremental useful results can be a burden. It's good because it keeps
you from moving a long way in a useless direction, but it can also
tremendously slow down progress toward long-term goals.
Earlier this year, in the US Virgin Islands, Marvin Minsky and his friends
had a private AI symposium on the topic of human-level intelligence. It was
written in up the June 2004 issue of the AI Magazine, and it seems to have
been a bit more interesting than the AAAI symposium that we just attended.
No real solutions were proposed, though; the focus was on Minsky's and
Sloman's architectures for human-level AI (e.g. Minsky's Emotion Machine
stuff).
One idea proposed by Minsky at that conference is something I disagree with
pretty radically. He says that until we understand human-level
intelligence, we should make our theories of mind as complex as possible,
rather than simplifying them -- for fear of leaving something out! This
reminds me of some of the mistakes we made at Webmind Inc. I believe our
approach to AI there was fundamentally sound, yet the theory underlying it
(not the philosophy of mind, but the intermediate level
computational-cog-sci theory) was too complex which led to a software system
that was too large and complex and hard to maintain and tune. Contra Minsky
and Webmind, in Novamente I've sought to create the simplest possible design
that accounts for all the diverse phenomena of mind on an emergent level.
Minsky is really trying to jam every aspect of the mind into his design on
the explicit level.
Another idea that came up at the Virgin Islands symposium was to create a
simulation world in which AI systems control agents that collectively try to
solve simple object-manipulation tasks. The prototype case is a bunch of
kids collaborating to build towers out of blocks. The idea was also raised
of making the simulation more realistic by making the block-building take
place in a simulated livingroom or restaurant or some such. I like this
line of thinking because it is closely related to the AGI-SIM simulation
world project that we're currently working on (an open-source sim-world to
be used for Novamente bue also by other projects).
-- Ben
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:49 MDT