Re: [opencog-dev] CMR Protocol [was: restructuring of OpenCog wiki site]

[Matt Chapman]

Hi Matt,

Your protocol makes a lot of sense to me, and it seems like it wouldn't be especially hard to implement a OpenCog Mind Agent to receive CMR messages and add them to the Atomspace, and another to broadcast CMR messages for newly created atoms. Is that something you've done or considered?

All the Best,

Matt Chapman

On Mon, May 7, 2012 at 4:39 PM, Matt Mahoney <mattma...@gmail.com> wrote:

First I want to thank Ben for bringing me up to date on the work that
the Hong Kong group is doing. Apparently they have their own separate
mailing list which I'm not on because I'm not actively doing any work
with them.

Which I guess raises the question of what I'm doing on this list. I'm
interested in AGI, like a lot of people, so I joined to mostly lurk
and learn more about OpenCog. My real job is doing basic and applied
research in data compression. I am currently at Dell as the result of
their acquisition of Ocarina, a small Silicon Valley company that
makes network storage optimizers using advanced compression
algorithms. I telecommute from Florida and I have a lot of flexibility
in both my hours and in my line of research. As you probably know,
data compression is an AI problem. More specifically, text compression
is equivalent to passing the Turing test, and general (algorithmic)
compression is equivalent to AIXI. My interest in OpenCog, as well as
some other AGI projects, is to see if there is any overlap with my
research.

It seems that the Hong Kong group is developing intelligent characters
for multiplayer video games. That's interesting, but doesn't really
fit in with my work. I also don't think it will lead directly to AGI,
although I think it will make a small contribution, just like many
other AI projects have done.

As I said, AGI is a hard problem. It does not help that no two people
can seem to agree on the solution. It seems that on all of the AI and
AGI lists that I have joined over the years that the quality of the
conversations has suffered as everyone who is doing real work has left
to go off in their own direction.

Those who have been here for awhile are familiar with my proposal.
http://mattmahoney.net/agi2.html

To summarize, AGI is made up of lots of narrow AI specialists, plus a
distributed message passing network to route questions and responses
to the right experts. To the user, it will look like a global message
pool where you can say anything you want and it goes to anyone who
cares. The messages that you see are ranked by your interests, which
can be determined from every message in the pool that you have ever
posted.

I believe that AGI has to be built on a global scale. The target is
not to make a machine that is smarter than a human, but a network of
machines that are collectively smarter than humanity. That requires a
decentralized global effort. In my design, both the experts and the
message routers are developed independently by people who may be
competing against each other. The driving economic force is
competition for your attention in a world where information has
negative value. Peers that don't deliver useful and valuable
information will be blocked. Peers that do a good job of holding your
interest will be of value to advertisers.

The protocol is not difficult. The main requirement is pairwise key
exchange and digital signatures to allow the development of reliable
reputation networks. Nor is it difficult to contribute to the
knowledge base. Any human knowledge will do, even posting pictures of
last night's drunken party. I worked out the scalability in my thesis.
Basically it requires O(n log n) storage and O(log n) message hops.
Peers would tend to specialize, which would further improve both space
and time efficiency. There are economic incentives to add redundancy
to make the network robust.

The hard part of the problem is routing messages. A message can be
natural language text, pictures, or video, tagged with the sender's
address, time stamp, and signature. The problem to be solved is
determining the mutual information between messages, so that when it
is high, each is routed to the address of the other. You need to be
able to match, e.g. questions to answers, or the names of people to
their photos.

This is a compression problem. It is also a problem that I think
OpenCog might solve, if we had sufficient test data, computing power,
and the right algorithms. It is not a critical step, however, as
others will be developing competing algorithms and users will decide
the winners. If people want to go off in their own directions, then
that's fine. It will make the system stronger.


-- Matt Mahoney, mattma...@gmail.com

--
You received this message because you are subscribed to the Google Groups "opencog" group.
To post to this group, send email to ope...@googlegroups.com.
To unsubscribe from this group, send email to opencog+u...@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/opencog?hl=en.

[Matt Mahoney]

On Mon, May 7, 2012 at 7:56 PM, Matt Chapman <ma...@ninjitsuweb.com> wrote:
> Hi Matt,
>
> Your protocol makes a lot of sense to me, and it seems like it wouldn't be
> especially hard to implement a OpenCog Mind Agent to receive CMR messages
> and add them to the Atomspace, and another to broadcast CMR messages for
> newly created atoms. Is that something you've done or considered?

I haven't implemented CMR. It really wouldn't be useful to people
until a lot of people were already using it. And of course, people
aren't going to use it if it's not useful.

It turns out that the internet is evolving in this direction already,
except that the protocols are much more complex. You have big, high
reputation peers like Google and Facebook communicating with many
other websites. And of course they use a lot of AI. I think that big,
high reputation peers would develop under CMR as well. You trust X,
and X says you can trust Y. We just have to make sure that untrusted
sites can't pretend to be X.

So in either case, you still have the hard problem of mutual
information. In Google, it's matching queries to web pages or images
or ads to your email. In Facebook, it's matching posts to your
interests based on your previous posts and likes. They also
experimented with automatically tagging faces for awhile.

You don't have to solve the whole problem. CMR works even if peers
only know how to match a small subset of messages. Other peers learn
what their neighbors specialize in by sampling their messages, and
reinforce that specialty by only sending related messages. You end up
with a routing network that looks like a partially hierarchical
ontology, but it is still fully connected.

One problem that CMR would solve is Google's monopoly on search. You
can't have n search engines because then you either have to increase
spider traffic by n, or increase index update delays by n in order to
maintain the same traffic level. With CMR, there is no central search
engine, censorship is harder, and the index is updated immediately
whenever something is posted. Google would still have the advantage of
a higher reputation than smaller websites.

In any case, I think that compression or prediction has interesting AI
applications, either with or without CMR.


-- Matt Mahoney, mattma...@gmail.com

[Matt Chapman]

On Mon, May 7, 2012 at 6:06 PM, Matt Mahoney <mattma...@gmail.com> wrote:

And of course, people
aren't going to use it if it's not useful.

That's where I might disagree. People might use it just because it's interesting and available. But they *can't* use it until it's implemented in *something*. Hence my suggestion of implementing it as a Mind Agent, not least because this is the OpenCog mailing list. Also, since you seem to have an interest in seeing OpenCog make practical progress toward useful applications, I thought this might be a possible contribution that would be interesting to you. 

Lastly, it seems to me that the intelligence of your system is actually in the filtering of messages, and your proposal is short on suggestions for how that might be implemented, but OpenCog provides a mechanism by way of Attention Values (accept Messages that would be have links generated (by PLN, or whatever) to atoms of with Attention Values above some threshold.)

[Ben Goertzel]

CMR seems an interesting idea, but it seems to require a huge amount
of both marketing and infrastructure work, quite separately from any
AI component...

Using OpenCog as a component of a CMR system could make sense, but
there would also need to be a lot of non-OpenCog work done...

So the OpenCog list seems a suboptimal place to recruit folks for CMR,
at this stage...

-- Ben G

> --
> You received this message because you are subscribed to the Google Groups
> "opencog" group.
> To post to this group, send email to ope...@googlegroups.com.
> To unsubscribe from this group, send email to
> opencog+u...@googlegroups.com.
> For more options, visit this group at
> http://groups.google.com/group/opencog?hl=en.

--
Ben Goertzel, PhD
http://goertzel.org

"My humanity is a constant self-overcoming" -- Friedrich Nietzsche

[Matt Mahoney]

On Mon, May 7, 2012 at 10:54 PM, Ben Goertzel <b...@goertzel.org> wrote:
> CMR seems an interesting idea, but it seems to require a huge amount
> of both marketing and infrastructure work, quite separately from any
> AI component...
>
> Using OpenCog as a component of a CMR system could  make sense, but
> there would also need to be a lot of non-OpenCog work done...
>
> So the OpenCog list seems a suboptimal place to recruit folks for CMR,
> at this stage...

That's not my goal. As I said, I have a real job doing data
compression research. Any work I might do with OpenCog has to relate
to that. There is a lot of overlap with AI, namely:

- natural language modeling, which is important for lossless text
compression, and
- human-understandable feature extraction from images and video, which
is important for lossy compression.

Now I *could* throw together a simple CMR implementation. It would
have a text window where you could post messages about anything, and
you would see a list of messages posted by other people, perhaps
ordered from newest to oldest. As I developed this, I could add more
intelligent ranking based on keyword matching, "like" and "spam"
buttons, and such. Others could download the open source software,
connect to it, and develop their own ranking algorithms.

But this is quite different from building AGI. To use an analogy, what
I did was like writing the specification for the first versions of the
HTML and HTTP protocols, as opposed to building the internet. I
wouldn't say there is a zero chance of success. The first version of
the Mosaic browser and NCSA web server was a 6 week effort by 1
person.

So you might wonder why I mentioned CMR, which I proposed 4 years ago
and never tried to implement. Maybe I didn't want to spend 6 weeks,
only to have it become a full time job. You are right that it would be
a lot of work. I estimate $1 quadrillion, or 20-30 years of global
effort. The only reason we would ever spend so much money is that we
now spend $70 trillion per year worldwide to pay people to do work
that machines could do if they were smarter. I don't expect that there
is any quick and simple solution to this problem. There is a good
reason why pioneers like Minsky and Kurzweil are not trying to build
AGI. Minsky will tell you: it's because intelligence isn't one thing;
it's solving lots of smaller problems that aren't very intelligent by
themselves. But I'm sure you know this.

More to the point. I don't think that building an "artificial toddler"
and later a self-improving "artificial scientist" is the way to solve
AGI. AGI is not an artificial person. I use my proposal to describe
what I think AGI will look like. Regardless of whether CMR is
implemented as an underlying protocol, we are going to have a smarter
Internet with a distributed mix of human and machine intelligence
contributing to a global knowledge pool. Nobody is going to be famous
for solving the "AGI problem", just like nobody created the internet.
We need to have realistic and achievable goals.


-- Matt Mahoney, mattma...@gmail.com