Hi Dan,

AI is a wonderful study. It has the grand goal of making philosophy an
empirical science. Long may it wave. However, it has not been
particularly useful in designing robots.
.
The problem of the brain can be stated in a lot of ways. Here are mine.
.
There are 10^15 neuron synapses in the 10^11 neurons of the brain. These
are - literally - astronomical numbers. This is equivalent to 1,000,000
gigabytes at 8-bits of resolution per synapse. Consider this a program
that is 1 exabyte long, for which we have no source code. We do not even
have a complete memory dump. There is no way to measure the 10^15
synapse values.
.
These synapses have an initial value when they are grown at birth and
after. We do not have access to these values. Next, the synapses change
as a result of learning = as a function of the sequence of nerve pulses.
We do not have access to the complete set of actual nerve pulses and
their sequence. A neural network develops as a function of the initial
synapse values and the sequence of training examples applied to it. We
have access to neither.
.
The operative assumption is that the brain evolved over a period of
millions of years. We *really* do not have access to the initial
conditions and sequence of training examples applied to the human
ancestors over this period of time. So, we do not know how the brain
evolved, other than in *extremely* gross outline.
.
The brain evolved to survive in its environment, and we do not know all
the critical things (small and large) that allowed the survivors to
survive. We do not have access to the examples that did not survive. So,
here is another area of variation that we cannot access.
.
There appear to be no "laws of the brain," in the sense of Newton's
laws. Not counting the centuries of philosophical speculation, it has
been 50+ years of intense effort, and no laws of the brain have been
found. My hypothesis is that the brain contains a huge collection of
evolved "hacks," each of which were successful in an evolutionary
survival sense. Instead of a simple machine with a few operating rules,
we have what amouts to a landscape of very many, randomly evolved
solutions to point problems. So there will be millions of laws, not a
few.
.
We tend to want to believe that the brain has simple rules. I ascribe
this to two things: the success of science in finding simple rules for a
great many things, beginning with mechanics and astronomy. It includes
genetics, chemistry, and a great many other successes. So, it was worth
a try.
.
The second thing that makes us guess that there will be simple rules is
language. We can all talk. It is the basis of most defitions of
intelligence. See Turing Test. Language has (some) simple rules:
grammar, syntax, etc. We all tend to believe that the thinking,
internally-talking part of our brain is the source of all brain
activity. All hail the conquering ego. However, language and thought
have a bandwidth of about 16 Hz. (See the book "The User Illusion" by
Tor Norretranders.) The raw "processing" bandwidth of the neurons at 100
Hz each is 10^15*10^2 = 10^17 = 100 exabytes/second. Most (!) of the
action is going on below the conscious level.
.
The Stanford DARPA challenge car used heuristics, not AI to win. It may
have used various pieces of AI related algorithms for specific problems,
but the reason it won was extensive hacking. They did a *lot* of
"mechanical debug" work in the field, discovering problems with specific
things in the desert, and they developed code to handle those problems.
This was the lesson from the first DARPA challenge. No car got beyond
the first 10 miles for various reasons that seemed simple and dumb in
the past tense. The real reason was little or no field testing. The
second time, all the cars spent a lot of time in the desert, trying to
make something work. So, the Stanford car path through the desert was
not a derivation from AI theory, it was a working set of hacks, like
allmost all successful robots..
So what? How do we build a robot, assuming that is our goal?
.
I believe that we need to reject AI as a robot design model.
.
I believe that we can make a full, just-like-in-the-movies (tm) robot
using object recognition and straight-forward engineering. I believe the
design model for the robot is not a synthetic human with independent
thinking, but a sheep dog that acts as an agent for the shepherd. The
sheep dog has autonomy (like a thermostat) without independence. And I
have spelled this out. Check my paper "Autonomy Without Independence"
from the second NASA Workshop on Radical Agent Concepts (WRAC), 2005.

Dave

--- In SeattleRobotics@yahoogroups.com, "dan michaels" <oric_dan@...>
wrote:
>
> --- In SeattleRobotics@yahoogroups.com, "David Wyland" dcwyland@
> wrote:
> >
>
> > My contention is that we are looking in the wrong place. If the
keys
> > were lost elsewhere, it does not matter how bright you make the
light
> > inthe lamp post. (I have my own ideas on where to look, and have
> > published them. This is another topic.) I believe the problem is
> simple
> > for complicated reasons, not the other way around.
> >
> > Dave
> >
>
>
> We only have one real existence proof of true intelligence, and that
> is the brain. In the old days, they thought the brain was mostly a few
> passive sensory input areas, followed by rather largish so-
> called "association" [learning] areas, ie essentially most of the rest
> of the brain. This idea apparently came down from Kant's concept of
> intelligence involving [passive] sensation and central [top-
> down] "understanding".
>
> Now, we know the brain is actually a hodge-podge of smallish special-
> purpose processors, produced by evolution tinkering with, and
> modifying, pre-existing structures. One of the best "structural"
> examples of this is that the ossicles of the inner ear of mammals
> evolved from the rearmost jawbone of reptiles, which evolved from the
> rearmost gill of bony fishes. In fact, this change from jawbone to
> ossicle is apparently still seen in kangaroos between when they are
> born and when they finally emerge from their mother's pouch.
>
> Back to the brain, about 40% of the cortex is devoted to vision, and
> which comprises greater than 30 separate special-purpose processing
> areas. Different areas compute various operations on the visual image,
> such as line-orientation detection, color-separation, binocular
> disparity, directional motion, form, texture, etc. A good book on this
> is Semir Zeki's 1992 book called A Vision of the Brain. See also
> here ...
>
>
http://defiant.ssc.uwo.ca/Jody_web/fMRI4Dummies/functional_brain_areas.
> htm
>
> IOW, whatever learning there is is preceded by massive amounts of
> special-purpose processing, and which is attuned to the specific
> problems to be solved, in the different modalities. [this is actually
> a form of functional decomposition]. I think we see that the Stanford
> Darpa guys pretty much followed this approach.
>