Re: Computer being developed modeled after human brain

On Apr 29, 10:31 pm, c...@xxxxxxxx (Curt Welch) wrote:
casey <jgkjca...@xxxxxxxxxxxx> wrote:
There is always a trade off between learned and
innate skills.


What do you mean by that? What trade off are you
talking about?

Write a tic tac toe program with innate actions and
one that learns. Which one uses the most resources
in time, energy and machinery?

Predators depend a lot on learned behaviours whereas
prey animals like sheep and horses do not. Sheep
do not have to hunt for food, lions do. There is
a payoff for adaptive behavior in a lion but not
in a sheep.

So there would be a general trend to favour innate
behaviors where possible and learned behaviors
only when there is an advantage in doing so.


Your idea of needing a module for "edge detection"
is exactly like suggesting you need to solve a
data compression problem by first writing a
run-length-encoding compression module.


If the "edge detection" algorithm is not useful for
a wide range of vision problems and some other
algorithm comes along that is more useful for a
wider range of vision problems then the current
edge detection algorithm will be replaced.

It is about evolving in working stages. You seem
to think we can jump from A to Z without any trials
and mistakes along the way.


run-length-encoding is a simple but very poor type
of compression where you look for repetitions of
characters, and replace them with an encoded
messages that includes the charger and the number
of times it was repeated. It works well if the
data you want to compress tends to have very long
strings of the same character, but fails to compress
any other type of data.


And if the environment is one of very long strings
of the same character then that compression algorithm
will do fine. If not it will mutate toward algorithms
that are more general in nature.


The best algorithms for compression are GENERIC
compression algorithms that work well for a very
wide range of different types of data. They are
not specific modules that only work for a very
limited type of data.


And do they use modules that are used in other types
of algorithms like using the math coprocessor?


The data processing required to create human and
animal like behavior is likewise just like that.
The very reason AI hasn't done better than it
has, is exactly because it's doing what you are
suggesting - using algorithms that are too
simplistic, and to narrow in the focus.


So are you suggesting the first neural systems out of
which the human brain evolved weren't simplistic and
narrow in focus to start with?

If your algorithm enables the robot to survive better
than my algorithm than your algorithm will have
reproductive success.


You can't create a strong compression system by
first doing run-length-encoding, and then adding
other such modules on top of that. You create
strong compression by throwing out those bad
algorithms that only work in very limited domains,
and replacing them with highly generic algorithms.


Sure I have no problem with that. The more generic
the edge algorithm the better. The human visual
system can see edges that none of my edge programs
can see.


Though there are edges in the world, most the
information in our visual field lives in the
continuous shades and the subtle ways that change
as objects move. We can recognize complex contours
not because of edges, but because of the subtle ways
the shading and colors change as they move.


The better edge detectors tend to require more
computational resources and so there has to be
a payoff.

We might take the crude edge detection I might use
as equivalent to the crude clump of light sensitive
cells that led to a complex eye.


The only way to create strong AI, is to create
generic algorithms for doing data extraction and
compression in a parallel real time information
network. If we try to hand-code simple ***, the
result will always be simple-***-AI, which we
have been building for 50 years now (***, it's
almost 60 years now - my how time flys...).


At this point in time simple *** as you call it is
all we have and is all the first neural networks
started with and they managed to evolve into this
generic data extraction parallel real time information
network you claim is required.


I understand why you personally may want to continue
hand-coding functions into a machine like edge
detection or whatever. That approach is something
you can actually build today.


If you know how to hard code a working eye there is
no point evolving it. The same applies for any useful
innate computational skills in a gene determined
neural network. Those skills may enable more and
more generic learning.


But what I don't grasp, is how it's not obvious to
you that human level AI will NEVER be created that way.
We know what the brain is, and it's not hard-coded modules.
It's a strong generic learning system all the way down to
individual pulses.


That is your opinion. I see it as 100% hard coded modules
at the bottom (even if fine turned and requiring input
data to develop) and 100% learned modules at the top. It
is modules at both ends. I use program modules that other
people learned how to make to build up bigger programs
that I am learning how to make. Their modules are not
visible in my programs and show no behavior without my
program but they are there and they are innate (that is
I didn't do the learning how to write them).


Your argument would probably be something like -
"well, removing duplicates is something that has to be
done, so why not make it the first module!".

The counter argument is simple. That's the WRONG MODULE,
and if you build the wrong module, you won't solve the
problem.


Yes I know you imagine a learning machine has to be built
out of a simple single learning module arranged in some
kind of network. I take an evolutionary view. You make the
"right" module and see where it takes you.


There's only one type of module that will solve AI, and
that's strong generic real time, parallel, high dimension,
reinforcement trained reaction network.


Those ideas go back to the early 50's and 60's.


I can continue writing like this and explaining more
specifics. But I've talked about all this many times
in the past, and it seems to just go over your head.


You make wild statements such as "the combinational
explosion problem has been solved" and it turns out
to be nothing more than ideas touted way back in the
50's and 60's such as Uttley's classification system.

Much of what you have written, apart from your temporal
emphasis, including what I have just edited out, I read
about years ago. I have no issue with those ideas just
that nothing practical has been built from those ideas.

I don't see any technology advancing anywhere in one
giant leap such as no flight to a Boeing 747 without
all the steps in between. Each stage is useful. This
is how it worked for biological control systems and
how it will work for technological systems.

Sure a simple edge detector may go the way of the
run-length-encoding compression module but so what?
Do you think the first neural system for processing
vision had an instant human level edge detector?

My hunch is that reinforcement learning is always
limited by the system in which it operates and only
advances as the system changes. I see your notion
of a generic can do anything system is a myth.

JC



.

Loading