Re: Computer being developed modeled after human brain

casey <jgkjcasey@xxxxxxxxxxxx> wrote:
On Apr 26, 9:05=A0pm, c...@xxxxxxxx (Curt Welch) wrote:

Oh, come one. Are you telling me if I digitize a movie
and feed it to your program, you can make the program
correctly identify every object, and every action, which
takes place in the movie?

I never said I could write a program that could correctly
identify every object, I wrote that it would not make a
mistake between an object and an action.

Well, by "identify" I didn't really mean give the correct English name for
each object and action. I really just meant more like correctly point it
out on the screen and point out when the same object, or same behavior,
showed up multiple times in the move.

Of course it must
first identify a collection of pixels as constituting an
object and then connect that to another collection of
pixels in the next frame to compute its actions. I don't
see, however, if that is done, how the object and action
would be confused with each other.

Because you don't grasp that detecting patterns in frames is not the same
problem as detecting objects, and detecting motion from frame to frame is
not the same thing as detecting object behavior.

What happens when the eyes of the robot moves like a human and jumps all
over the place constantly? What happens when the head moves one way, as
the eyes move another, and the robot is riding a bike, on a moving train?
Everything is is in motion, but yet, how does it associate the various
motions correctly with the thing doing the behavior? Which part of the
motions are an indication that the eyes are moving? Which part of the
motions are an indication that train is moving?

Correctly parsing this type of data is a whole world of difference from
parsing a static scene of stacked blocks when the robot is standing still
and the blocks are standing still. It's even more complex when different
parts of the object are moving in different ways.

A human looks at that and sees one object - a rotating
Rubix cube. How does your software parse all those
changing pixels as "one cube" without making the mistake
of seeing it as many different moving color blobs? How
does it correctly find the object in all the motion?

Are you interested in working out possible solutions to
that problem or are you going to leave it all to your
magic net to do it for you?

I'm interested in working out how to build those magic nets.

TD-Gammon uses "magic nets" as you call them to make all the decisions. It
works better than any hand-coded module for backgammon created by the same
author or any other author. It worked by programing itself to create the
right algorithm. But once it was done programming itself, the entire
program was nothing more than than the list of weights for the neural
network. Do you thin that Tesauro could look at those numbers and
"understand" what the code was doing? I'd be willing to bet that for the
most part, he couldn't even begin to read the code. Can he for example
explain to us why one weight was .34568 and not .34569? Not likely.

What you are suggesting, is that if we were to sit down and try to
hand-code an object recognition algorithm we would learn how to program
those numbers better? No way. I've been saying this for over 5 years now,
and you still don't seem to get it. YOU ARE WORKING ON THE WRONG PROBLEM.
YOU ARE BUILDING THE WRONG TYPE OF MODULE.

Though such work can and does produce all sorts of interesting new
machines, it will NEVER produce real AI because that's just not what real
AI is.

Yes, I'm very interested in solving real AI, and every since I came to this
group, I've been looking for people who are working on it - and sadly, I've
not yet found any. So instead of getting to work with people to solve the
problem, I spend all my time here trying to get more people to realize what
the problem is.

And inversely, given a still photo of an action shot -
a player hitting a ball, how does the program correctly
identify the behaviors when nothing is moving?

Interesting questions Curt and in an area of great interest
to me. However until your magic net comes on line we will
have to figure out some solutions ourselves and see if your
magic net comes up with similar solutions.

If your "other solution" is to hand design some modules like edge
detection, then you are wasting your time. The only solution that will
solve this requires a generic information compression algorithms that
compress information from the temporal domain to the spatial domain by
making predictions about what will happen next. And ones that are then
able to be trained by reinforcement. If you are not working on the
solution to that problem, you simply aren't working on real AI, you might
as well be writing a word processor because they have about as much in
common with real AI as what you are suggesting.

Again, I have no issue with people wanting to explore that path to see what
sort of hand-designed behaviors that can create - people even get rich at
times coming up with new algorithms like that. But it's just not real AI.

Computer systems that attempt to mimic this more complex
process of prediction by looking at areas of similar
pixel data, and then trying to track how that area moves
from frame to frame, will totally fail to see the
rotating Rubix cube as one object, and will totally fail
to see a video of a static picture of a Rubix cube
dropped from a hand (the cube is in mid air underneath
the open hand), as motion.

Computer vision systems can already track moving objects
such as pedestrians or thrown balls. As for a single frame
of a falling object, even you could not determine its motion
except to the same extent a program could by determining
it was not supported and was probably falling under the
force of gravity. Or perhaps the behavior of a person by
the frozen "walking" configuration of their bodies?

Evolution of course had no ultimate goal

Yes it does. Survival is the ultimate goal that drives all of evolution.

but goals did
evolve in animals because they provided a reproductive
advantage.

Because they were good sub-goals of the ultimate goal of surviving.

If we wish to bring about some desired goal
such as AI we might try and duplicate evolution

Yes, if we had a 100 million years to spare and a huge ass computer to run
the simulation on, we might consider that.

or we
might do what we have done with artificial flight and
figure out a solution. That way we are part of the AI
learning loop which gives us some idea what it is we do
that needs to be replaced in an self contained AI loop.

Yes, but if you work on the wrong solution, you will never get your
software "off the ground". Just look at all the things that were tried for
powered flight and how utterly stupid so many of them look today with our
advantage of hindsight.

Whether you understand it or not, witting an edge detection module as a
step to creating an AI vision system, is as stupid as trying to build a
plane with flapping wings.

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

Relevant Pages

  • Re: SR and LET Are Incomplete Aether Theories
    ... >> Motion wrt the rest frame of the ether is called absolute motion. ... The observer is the one who is doing the moving. ...
    (sci.physics.relativity)
  • Re: A nice puzzle: How fast does the E field move?
    ... >> Now, with C moving at 0.5c, an event on ship A and an event on ship C ... > C's frame, if they are to be simultaneous in the stationary frame. ... Motion is along the X axis. ... You always get to have a plane of simultaneity containing any ...
    (sci.physics.relativity)
  • estimating motion in pixels/frame
    ... i have a sequence of images which represent moving text and ... my project requires me to estimate the motion in pixels per ... frame. ...
    (comp.soft-sys.matlab)
  • Re: mass increase due to speed
    ... The ratio of gravitational to inertial mass is fixed, ... it must also gravitate more. ... Suppose a body is moving along the x axis at 3-velocity V, ... All mass measurements in your frame are with respect to other masses ...
    (sci.physics.relativity)
  • Re: Electrons prevent Cosmic Ray Black Holes
    ... 'K' is the stationary system of coordinates, 'k' is the moving frame. ... between the two cars is falling at the rate of 100 kph, all speeds ...
    (sci.physics.relativity)
Loading