Re: Economy of ideas
- From: curt@xxxxxxxx (Curt Welch)
- Date: 09 Sep 2006 23:07:10 GMT
"Lars" <LarsFiedler@xxxxxx> wrote:
Economy-Problem
---------------------------
1. Ideas are representations of perceptions. An idea is represented by
multiple neurons.
2. Human thinking is based on assoziativity of ideas.
There must be a biological mechanism that adjusts the number of active
ideas at a time. Has anyone heard of such a biological mechanism?
I don't know much about brain physiology, but sure, it's well known that
there are systems at work regulating brain activity. When they fail, you
get events like an epileptic seizures. I don't know how much is know about
the different regulation systems, but the brain must have them. If neurons
fan-out to 10,000 other neurons, there must be systems that regulate
overall activity. If there were not, one of two things would happen.
Either stimulus signals would die out quickly in the network, and you would
end up with the brain producing no outputs at all, or stimulus signals
would quickly spread and cause every neuron in the brain to fire in near
unison. Neither of these cases would be good.
Has
anyone been confronted with this problem in AI-programming? (Note that
A,B,C are not single neurons but representations of ideas.)
Yes. All the time. It's one of the prime problems I've run into when
trying to train neural networks with reinforcement learning. If you have
10 sensory inputs, and one input fans-out too aggressively in the network,
it tends to dominate the network and cut off the activity of the other
sensory signals. And likewise, a sensory signal which doesn't survive
multiple levels of a network, is lost, meaning the output created by the
network can't be a function of that sensory signal. I've played with many
different techniques to regulate these issues but they all map to your
concept of "economy of ideas".
At a higher level, the general problem of trying to create intelligent
machine behavior has always been a problem of making the machine select
what to do next. When the "what to do" is a long sequence, the machine
must pick some behavior goal and stick to it. It can't do two things at
once - like standing up, and sitting down, at the same time. If it tries
to do both at once, the machine is likely to just fall over and not end up
standing, or sitting. All behavior decisions requires that the machine has
the ability to create an economy of behavior (to steal your use of the
word) in order that it never wastes time trying to do two conflicting
things at once. We only have one mouth and we can only speak one word at a
time. We only have one right hand and can only grab one thing at a time.
etc. All operation of a machine like a human body requires solving the
economy of behavior problem.
If you are attempting to control a complex machine with a network that is
inherently parallel (a neural network), then you have an instant economy of
behavior problem to solve when you get 10,000 neurons all trying to tell a
single arm to do something different. One way or another, that problem has
to be solved.
We see a solution in action in the brain for example in visual illusions.
The brain tends to lock on to one answer or another as to what we are
looking at. If you look at a wire figure drawing of a cube our brain can
make it flip between two different views. But, once flipped, the brain
tends to lock onto that view until it is forced to flip to the other view.
This is the system that creates economy of ideas in action. People talk
about this working like strong attractors in a chaotic system.
--
Curt Welch http://CurtWelch.Com/
curt@xxxxxxxx http://NewsReader.Com/
.
- References:
- Economy of ideas
- From: Lars
- Economy of ideas
- Prev by Date: Re: Bayesian Inference Engine
- Next by Date: Re: Economy of ideas
- Previous by thread: Re: Economy of ideas
- Next by thread: Re: Economy of ideas
- Index(es):
Relevant Pages
|
