Re: No easy road to AI

casey <jgkjcasey@xxxxxxxxxxxx> wrote:
On Dec 12, 3:54=A0pm, c...@xxxxxxxx (Curt Welch) wrote:
casey <jgkjca...@xxxxxxxxxxxx> wrote:
The advantage a circuit has over a look up table is it doesn't
have to insert data to look up the response, it simply computes
the response.


The difference here is not important to our abstract understanding
of what type of process we are dealing with. It's only important
in terms of the implementation details we choose to use when
we build it.

Perhaps you could show me how you would implement a look up table
to transform an image into a list of blobs?

Well, that's a good question. Nope, I don't know hand-create such a
function. I'm fairly sure it could be done however and studying that
problem would be one of many approaches to gaining a better understanding
of this problem and how to solve it.

Pictures however are not temporal data. They are purely spatial data. As
such, wasting a lot of time solving only that problem would not be good
because when you are done, you would have an algorithm that is unworkable
for the domain we need to solve - which is in the temporal domain.

Humans have networks that decode 2D visual data. But I don't believe that
network was formed either by evolution, or by exposure to lots of pictures.
It was formed by exposure to the temporal data that comes from continuous
interaction with a 3D world. When we look a picture, we "see" the 3D
objects it represents (assuming it's a photo or drawing that does have the
correct shapes to represent 3D objects). But that ability I believe didn't
come from looking at pictures. It came from looking at real 3D objects and
how they change over time when we move relative to the object. How that
data changes is constrained by what happens when a 3D object is mapped to a
2D representation in the eye. It's from those constraints that the design
of the network evolves from as our visual cortex wires itself. And once
wired to correctly identify expected correlations in real images, the eye
can also use that same transform to identify 3D objects in a picture.

It [Curt's theory] works because a combination, even when very
wrong, can be tested for "closeness" to being right.

Major breakthrough then? Shame you can't get anyone to understand it.

It's not a major breakthrough. It's basically how lots of RL algorithms
work - even though they tend to use a similar technique in a different way.
It's basically how TD-Gammon works. TD-Gammon faces the same sort of
combinatorial explosion problem in terms of the number of games it would
have to play to get enough experience but yet it solves it - which is why
TD-Gammon is such an interesting RL algorithm. It shows that the problem
you believe can't be solved, can be solved.

This clustering of features that gets adjusted I talked about is also known
as an abstraction of the sensory data. It's an abstract feature of the
data. TD-Gammon was not good at creating optimal abstract features - at
adjusting the way the data was clustered to revel the maximal amount of
information about how to make a move. It worked isntead, because the human
that designed it used his own intelligence which used is how experience in
designing Backgammon games to pick a clustering that he thought would work
well in Backgammon - and he was right.

So TD-Gammon can learn using abstractions, but what it fails at, is doing a
good job of finding it's own abstractions. To solve the general problem,
we have to build a machine that can not only abstract by transforming the
data, but which can also identify, and converge on, GOOD abstractions for
whatever problem it is working on. And the approach I'm talking about an
attempt to do that.

The point of all this posting is because there are unsolved implementation
questions here that I would like to find solutions for. I've been working
on this problem, ALONE for decades as a fun hobby. I came to c.a.p. years
ago hoping to find someone to brainstorm with about how to solve this
problem.

Instead of finding anyone to brainstorm with, I've found lots of people
that don't even understand what the problem is that needs to be solved. So
before I can get anyone to do some interesting design brainstorming with, I
have to educate the world (it feels like at times) about what the problem
is. Tough I keep trying, it's not yet worked. :) So I continue to
brainstorm about how to tackle this problem on my own (often in long posts
that might otherwise appear to be me trying to describe something to
people).

I see the combinational explosion problem as unsolvable and can only
be circumvented by simplification and the invention of heuristics.

Your "circumvented by simplification" is what drives me up the wall. Most
of your suggestions along that direction are obviously impossible. What
you are mostly doing there is denying the problem exists and pretending
it's solved in ways that it can't be solved and then you just use those
rationalizations to not work on the real problem (not even think about how
to solve it becuase you have convinced yourself it's unsolvable).

If "invention of heuristics" means "invent an algorithm that solves it"
then that's exactly what I'm working on and searching for. If it means,
"invent a trick so we don't have to solve it", then I'm back to being
driven up a wall. :)

The simple black-box problem I work on is the problem of taking multiple
parallel sensory inputs to the box and have the box produce multiple
parallel output signals. The box has a reward signal input, and must
adjust how it those outputs are computed from the inputs to make the
reward signal get higher over time - with no a prior knowledge of what
environment it's interacting with outside the box.

This black box problem is not intended to be the complete and full picture
of what the solution to AI will look like. But it's the key "magic box"
that's missing from all AI work currently. The complete solution might
includes lots of custom modules outside the box, and might include the
algorithm inside the box optimized to the problem. But until someone
solves this specific black box problem, no one is going to make any real
program in AI, or in brain research. Understanding how to solve this
specific black box problem, is the foundation of understanding how the
brain creates intelligent behaviors and how we are going to build machines
to duplicate that intelligence.

DO you grasp that no one has a workable algorithm to put in that box to
solve that class of problems? The fact that no one yet has such an
algorithm is why AI sill hasn't been solved in 60 years. The fact that so
few people are working on this problem is why progress in AI is moving so
slowly.

This generic learning black box problem is the only problem I've been
interested in for the past 30 or so years in AI and it's the only puzzle of
AI I'm going to be interested in until it's solved (or until I die).

--
Curt Welch http://CurtWelch.Com/
curt@xxxxxxxx http://NewsReader.Com/
.

Relevant Pages

  • Re: image normalisation
    ... Sounds like you need to look into histogram matching techniques (good ... does the brightness always come from some fixed, ... the pictures were taken with the light from down-right, ... threshold value in the algorithm it worked. ...
    (sci.image.processing)
  • Re: image normalisation
    ... tried the pictures they tried on. ... to admit, that his algorithm may might work, but only if he would ... does the brightness always come from some fixed, ... something like if i have one picture as a template, normalise the ...
    (sci.image.processing)
  • Re: Where is behavior AI now?
    ... regards going past the basics of intelligence. ... It's like documenting the flight path of a bird, without any understanding ... This is the same missing piece which has been missing for over 50 ... It's like trying to reverse engineer an encryption algorithm. ...
    (comp.robotics.misc)
  • Re: image normalisation
    ... i saw a dr. work from some guy about eye extraction. ... tried the pictures they tried on. ... to admit, that his algorithm may might work, but only if he would ... does the brightness always come from some fixed, ...
    (sci.image.processing)
  • Re: algorithm to sum run time of simultaenous jobs?
    ... jobs are a match for the size of the objects you're trying to fit into ... the 5 bins? ... then my understanding is that there is no algorithm outside of ...
    (rec.puzzles)