From:	Robert Watkins <robertdw@*******.NET.AU>
Subject:	Re: Neural nets 101
Date:	Fri, 6 Dec 1996 11:34:02 +1100

>Lets say this worked well, and the engine does the right thing. If the
>final answer
>was pretty "Yes" like, then every connection that said "yes" to
the output
>will get a
>tiny little boost in strength, and every connection that said "no" will get
>ever-so-slightly weakened. Conversely, if the engine doesn't do the right
>thing,
>every connection that said "yes" will be weakened and the "no"s
will be
>strengthened. You then go back and do the same thing with the connections
>between the inputs and the intermediates (gets trickier, but the concept
>is exactly
>the same). Even starting with random initial connection strengths, after
>enough
>training the net does the right thing. And once its learned the right
>thing, it can
>gradually learn as conditions change, as the engine wears out for example.

The other catch is that it's a probability thing... The chance still
exists in a neural net for the "No"s to get a series of successive
answers, which can make them the dominant choice. The most likely
candidate for this is if there is a sudden change in the conditions, as
opposed to a gradual change. The bigger you make your network, the less
likely this is, but the more complex it is. The more flexible you make
your value rules, the less likely this is, but the more complex it is.
For anything worth doing, you need very flexible rules, and a large
network. The resulting complexity is hellacious. The mammalian brain is
the end product of a billion or so years of evolution, and it still
breaks down.


--
Robert Watkins robertdw@*******.com.au
Real Programmers never work 9 to 5. If any real programmers
are around at 9 am, it's because they were up all night.