DARPA's Fun LoL project

[John F Sowa]

DARPA is starting a new project on trying to determine the
Fundamental Limits of Learning:

http://www.darpa.mil/news-events/2015-05-26

Some excerpts below.

My prediction: This project will be as successful as George Bush's
project, No Child Left Behind. The result will be a bunch of programs
for which the teachers (the programmers) learn how to make the pupils
(their programs) pass the test.

Fundamental principle: Children don't learn from examples. They learn
by doing, playing, and interacting with their parents and peers in as
many different ways as possible.

Example: If you have a Chinese nanny for your child, who talks and
plays with the child for X hours per day, the child will learn Chinese.

But if you put your child in front of a TV tuned to a Chinese station
for X hours a day, the child will learn nothing.

John

PS: When the Fun LoL project is canceled 5 years from now,
remember that you heard it here first.
______________________________________________________________________

http://www.darpa.mil/news-events/2015-05-26

DARPA seeks mathematical framework to characterize fundamental limits of
learning

What are the number of examples necessary for training to achieve a
given accuracy performance? (e.g., Would a training set with fewer than
the 30 million moves that programmers provided to this year’s winning
machine have sufficed to beat a Go grand champion? How do you know?)

How “efficient” is a given learning algorithm for a given problem?

What are the potential gains possible due to the statistical structure
of the model generating the data?

[Obrst, Leo J.]

John, I understood this program to be about mathematical limits/foundations of machine learning in AI.

Thanks,
Leo

>--
>All contributions to this forum by its members are made under an open content
>license, open publication license, open source or free software license. Unless
>otherwise specified, all Ontolog Forum content shall be subject to the Creative
>Commons CC-BY-SA 4.0 License or its successors.
>---
>You received this message because you are subscribed to the Google Groups
>"ontolog-forum" group.
>To unsubscribe from this group and stop receiving emails from it, send an email
>to ontolog-foru...@googlegroups.com.
>To post to this group, send email to ontolo...@googlegroups.com.
>Visit this group at https://groups.google.com/group/ontolog-forum.
>To view this discussion on the web visit
>https://groups.google.com/d/msgid/ontolog-forum/d4925838-5b50-d561-
>ae91-06e28522b935%40bestweb.net.
>For more options, visit https://groups.google.com/d/optout.

[Gary Berg-Cross]

John,

I guess that one might take a step in a fruitful direction by adding some contextual setting and embedding in the real world for the learning.

So could have developmental robots with some degree of an environment to study the "learning efficiency" considering such things as  learning through exploration and experimentation,  self motivation and discovery, human–robot interaction etc.  The type of thing discussed a while ago in "

Bringing up robot: Fundamental mechanisms for creating a self-motivated, self-organizing...

​"​

Gary Berg-Cross, Ph.D.  

gberg...@gmail.com     

​​

​

http://ontolog.cim3.net/cgi-bin/wiki.pl?GaryBergCross

Member, Ontolog Board of Trustees

Independent Consultant

Potomac, MD

240-426-0770

[Simon Spero]

On Jun 2, 2016 10:29 AM, "Obrst, Leo J." <lob...@mitre.org> wrote:
>
> John, I understood this program to be about mathematical limits/foundations of machine learning in AI.

See e.g. The SEP article on Formal Learning Theory :
http://plato.stanford.edu/entries/learning-formal/

For misapplications of Formal Learning Theory  see Noam Chomsky.   ( http://norvig.com/chomsky.html )

[John F Sowa]

Leo, Gary, Rich, and Simon,

Machine learning at a human level has been the Holy Grail of AI
research for over half a century. I doubt that the DARPA challenge
will magically stimulate researchers to find it. If anybody does
make a breakthrough during the next 5 years, it will be based on
what they had already been doing without the DARPA incentive.

Leo

> I understood this program to be about mathematical limits/foundations
> of machine learning in AI.

Yes. And that's why its chances of doing anything useful are slim.
The example they give is Shannon's discovery of the basic formula
for information theory. But that analogy is misleading:

1. DARPA challenges are useful for developing some kind of application,
but they are unlikely to discover fundamental formulas.

2. As Shannon himself said, his formula has nothing to do with meaning.
But meaning is central to human and animal learning.

3. Conditions #1 and #2 imply that a winning solution is likely to be
a special-purpose method for passing the test. (That was the fatal
flaw in the "No child left behind" program.)

GBC

> I guess that one might take a step in a fruitful direction by adding
> some contextual setting and embedding in the real world for the
> learning.

Yes. They are called applications. Practical applications have
often been the best stimulus for major innovations in AI.

GBC

> So could have developmental robots with some degree of an environment
> to study the "learning efficiency" considering such things as learning
> through exploration and experimentation, self motivation and
> discovery, human–robot interaction etc.

That would be better. But the critical issue is having a strong
incentive, such as survival in a hostile environment.

RC
> This is the PDF of the paper cited by GBC:

> "Bringing up robot: Fundamental mechanisms for creating
> a self-motivated, self-organizing...​"​
>

> http://167.206.19.12/~jmarshall/papers/cbs05.pdf

Thanks for the URL. But the authors made a bad assumption on page 1:
> We believe that a significant pitfall exists in both the top-down
> and bottom-up task-oriented robot design methodologies: inherent
> _anthropomorphic bias_. This bias refers to the design of prespecified
> robot tasks...

The most intelligent birds and mammals spend the longest time learning
from their parents. Without that "parental bias", they're helpless.

There is nothing wrong with parental-morphic training of the basics.
Hunger and fear will provide more than enough "self motivation" for
learning the open-ended variety of methods for achieving the goals.
It's irrelevant whether the "parents" are or are not anthropoids.

SS

> SEP article on Formal Learning Theory :
> http://plato.stanford.edu/entries/learning-formal/
>
> For misapplications of Formal Learning Theory see Noam Chomsky.

> http://norvig.com/chomsky.html

Those references discuss the current state of the art. But the goal
of the DARPA challenge is to make a major breakthrough.

John

[Obrst, Leo J.]

I think we probably all have opinions on the grander vision/epistemology behind these questions, as Simon indicates. Often it is the same rationalist vs. empiricist, Apollonian vs. Dionysian, neat vs. scruffy, symbolic vs. stochastic, theory vs. practice, science vs. engineering, explanation vs. description, etc., stances from time immemorial. Name your bifurcation.

 

Thanks,

Leo

 

--

All contributions to this forum by its members are made under an open content license, open publication license, open source or free software license. Unless otherwise specified, all Ontolog Forum content shall be subject to the Creative Commons CC-BY-SA 4.0 License or its successors.
---
You received this message because you are subscribed to the Google Groups "ontolog-forum" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ontolog-foru...@googlegroups.com.
To post to this group, send email to ontolo...@googlegroups.com.
Visit this group at https://groups.google.com/group/ontolog-forum.

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/CADE8KM6cAeP%2BEnuSwAEC5vQzY4KbgSua3CzGkP-DnDu9CnO_sg%40mail.gmail.com.

[Obrst, Leo J.]

Many here poof meaning. Or place meaning, ala late Wittgenstein, in contextual correspondences only, much like distributional semantics in NLP does. Find the contexts where words occur / co-occur, and you've found the meaning. If you buy into that view, then the current machine learning / deep learning is right up your alley.

But I think both notions are at play: the typed, i.e., described and logically formalized meaning (more or less) and the token co-occurrence meaning. Poetry, for example, uses the substrate layers (vast linguistic correspondences) to great effect.

Thanks,
Leo

>-----Original Message-----
>From: ontolo...@googlegroups.com [mailto:ontolog-
>fo...@googlegroups.com] On Behalf Of John F Sowa

>--
>All contributions to this forum by its members are made under an open content
>license, open publication license, open source or free software license. Unless
>otherwise specified, all Ontolog Forum content shall be subject to the Creative
>Commons CC-BY-SA 4.0 License or its successors.
>---
>You received this message because you are subscribed to the Google Groups
>"ontolog-forum" group.
>To unsubscribe from this group and stop receiving emails from it, send an email
>to ontolog-foru...@googlegroups.com.
>To post to this group, send email to ontolo...@googlegroups.com.
>Visit this group at https://groups.google.com/group/ontolog-forum.
>To view this discussion on the web visit

>https://groups.google.com/d/msgid/ontolog-forum/7a27a17f-6c05-df9f-7fd5-
>4cbd72f60960%40bestweb.net.

[Nadin, Mihai]

Leo J. Obrst and the rest:  

 

Because you asked: My bifurcation -- Reaction vs. Anticipation. The first describes physical interaction. It is the quantity/number domain. Nomothetic: subject to law. Deterministic. Anticipation describes interaction pertinent to living processes. It is the meaning domain. Idiographic: Gestalt. Non-deterministic.

 

 

Mihai Nadin

www.nadin.ws

www.anteinstitute.org

[Gary Berg-Cross]

So OK, no surprise there is a

​ year old ​

video showing Google's

​ ​

DeepMind

​ ​

using

​ ​

Deep Q-learning

​ ​

to master

​ ​

Atari Breakout

​ game.​

https://www.youtube.com/watch?v=V1eYniJ0Rnk

​ 

It is one of those​ examples that used a very constrained game problems and can be learned without having to master semiotics or the interactions of many factors such as in the real world. In other words a game.

Gary Berg-Cross, Ph.D.  

gberg...@gmail.com     

​​

​

http://ontolog.cim3.net/cgi-bin/wiki.pl?GaryBergCross

Member, Ontolog Board of Trustees

Independent Consultant

Potomac, MD

240-426-0770

--
All contributions to this forum by its members are made under an open content license, open publication license, open source or free software license. Unless otherwise specified, all Ontolog Forum content shall be subject to the Creative Commons CC-BY-SA 4.0 License or its successors.
--- You received this message because you are subscribed to the Google Groups "ontolog-forum" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ontolog-foru...@googlegroups.com.
To post to this group, send email to ontolo...@googlegroups.com.
Visit this group at https://groups.google.com/group/ontolog-forum.

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/7a27a17f-6c05-df9f-7fd5-4cbd72f60960%40bestweb.net.

[Ed - 0x1b, Inc.]

original article has been removed 404/451/bitrot
http://www.darpa.mil/news-events/2015-05-26
see
https://web.archive.org/web/20160601100002/http://www.darpa.mil/news-events/2015-05-26

is it April again? ;)

> https://groups.google.com/d/msgid/ontolog-forum/CAMhe4f2iL4g%3DtZ-XCEsY%2B6w9DkX5DDnw8dnOk%3D8Y3M4q%2BwA-Pg%40mail.gmail.com.

[John F Sowa]

On 6/6/2016 2:51 PM, Gary Berg-Cross wrote:
> It is one of those​ examples that used a very constrained game problems
> and can be learned without having to master semiotics or
> the interactions of many factors such as in the real world. In other
> words a game.

I discussed that example in slides 26 to 28 of the following:
http://www.jfsowa.com/talks/nlu.pdf

Note that it does well on some Atari games. But it is not so good
on Space Invaders, which requires a modest amount of strategy --
and Deep Neural Nets (DNNs), by themselves, are not so good in
learning long-term strategy.

So Google's AlphaGo group trained a DNN to beat the world Go champion.
And Go is certainly a game that requires some of the longest of long-
term strategy. See slides 28 to 30.

But some points to ponder:

1. AlphaGo is a hybrid system:

a) It uses a DNN to learn how to estimate the value of a position.

b) For long-term strategy, it uses a version of Monte-Carlo search,
which is commonly used for Go systems that do not use DNNs.

2. And the amount of training that AlphaGo required to beat the world
Go champion required a supercomputer that enabled it to learn from
30 million Go positions.

3. No human could play or analyze that many positions. A human who
took 1 minute per move would have to play Go for 40 hours per week
for 240 years to learn from that many moves.

See the other slides of nlu.pdf for more comments on related issues.
Slide 2, by the way, has the URL for the video of my talk.

John

[John Bottoms]

On 6/6/2016 8:44 PM, John F Sowa wrote:

On 6/6/2016 2:51 PM, Gary Berg-Cross wrote:

It is one of those​ examples that used a very constrained game problems
and can be learned without having to master semiotics or
the interactions of many factors such as in the real world. In other
words a game.

One of the interesting aspects of games that often goes unnoticed is the naming of moves. We do it intuitively. There is the "Statue of Liberty" pass, the "Sicilian defense" and the "Hat Trick". But we don't name moves in other games. Maybe there are names for moves among checker players that we don't hear about. But most of the moves in an Atari game are descriptions of strategies. There appear to be certain constraints in naming processes that go unspoken. Any onomastic professionals here?

-John B

[Gary Berg-Cross]

​>

One of the interesting aspects of games that often goes unnoticed is the naming of moves. We do it intuitively. There is the "Statue of Liberty" pass, the "Sicilian defense" and the "Hat Trick". 

​

Self organizing maps (SOMs) do things like this chunking strategy by looking down at patterns of firing in lower "neural nets."  The SOM then represents behavioral sequences as chunks for a wide variety of simple tasks

​.​

Gary Berg-Cross, Ph.D.  

gberg...@gmail.com     

​​

​

http://ontolog.cim3.net/cgi-bin/wiki.pl?GaryBergCross

Member, Ontolog Board of Trustees

Independent Consultant

Potomac, MD

240-426-0770

--
All contributions to this forum by its members are made under an open content license, open publication license, open source or free software license. Unless otherwise specified, all Ontolog Forum content shall be subject to the Creative Commons CC-BY-SA 4.0 License or its successors.
---
You received this message because you are subscribed to the Google Groups "ontolog-forum" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ontolog-foru...@googlegroups.com.
To post to this group, send email to ontolo...@googlegroups.com.
Visit this group at https://groups.google.com/group/ontolog-forum.

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/1c2599f2-09a2-b5a0-b86c-f202b1f4c2bd%40firststarsystems.com.

[Obrst, Leo J.]

I think more generally it’s a kind of compiling that humans do, not just with games. Some are not necessarily “named”, but just accessed.

 

Thanks,

Leo

 

From: ontolo...@googlegroups.com [mailto:ontolo...@googlegroups.com] On Behalf Of Gary Berg-Cross

Sent: Tuesday, June 07, 2016 8:32 AM
To: ontolog-forum <ontolo...@googlegroups.com>

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/CAMhe4f2A%3DtomFkT%3D-aMZinjV2anNyQ%3D%3DqkPPsp0HdSdC8RW%2BXA%40mail.gmail.com.