toward a generic learning algorithm

[casey]

The reason "seeing" is so easy is because it is done for you
not because the problem is easy. That is why we can program
a computer to be a champion chess player but not a champion
real time visual problem solver. It also involves reverse
optics which has an infinite number of solutions. You can
start with a 3D simulated world data base and generate a 2D
array of colors on a computer screen but you cannot take the
2D array of colors and convert them back to a 3D world data
base without making some major assumptions which if wrong
will produce visual illusions. It is a case of reverse optics
which is like working out what two numbers were multiplied
together to produce a particular number vs. optics which is
being given two numbers and working out their product.

Essentially Curt's notion of a generic learning machine is
equivalent to evolving a brain. Real brains started with
nets simple enough to have come about by chance. Over
millions of years additions and modifications were made
and tested in small incremental working stages. Curt's
assumption is that the human brain hit upon some generic
learning module that has/could replace all the hard won
specialists of yesteryear and he hopes to hit upon an
equivalent machine by trying to figure out what are the
requirements for such a machine.

My contention is that the problems we *actually* solve are
constrained by a special purpose world not a generic world
of anything can happen and we already have special purpose
compact material efficient solutions honed by evolution
over millions of years for problems that have been there
since year dot. These are not general purpose nets as such
although a useful module may be duplicated millions of
times to serve different sensory-motor needs and seeing
them all together may give the impression of a generic net.
That is what I believe you see in the neocortex, cerebellum
and so on all the way down to the reticular system in the
spinal cord. Much like the network of memory units in a
computer system. By themselves they are useless.

Now Curt wrote, it's *what* a human has the power to _learn_
that's amazing. So what does the human learn that other
animals are unable to learn? Is it because they don't have
a generic learning system?

JC