Path: utzoo!utgpu!news-server.csri.toronto.edu!neuron.ai.toronto.edu!neat.cs.toronto.edu!radford
Newsgroups: comp.ai.neural-nets
From: radford@ai.toronto.edu (Radford Neal)
Subject: Re: references help
Message-ID: <90Oct17.163038edt.310@neuron.ai.toronto.edu>
Organization: Department of Computer Science, University of Toronto
References: <Qb6rXii00VIF43nUVY@andrew.cmu.edu>
Date: 17 Oct 90 20:30:58 GMT
Lines: 21

In article <Qb6rXii00VIF43nUVY@andrew.cmu.edu> zl03+@andrew.cmu.edu (Zoonky L. Lee) writes:

>Do anyone have any references for the differences of using raw data and
>statistics as input in BP?

If I've interpreted your question correctly, the following paper is
relevant:

    Hopfield, J. J. (1987) Learning algorithms and probability
    distributions in feed-forward and feed-back networks, _Proceedings 
    of the National Academy of Sciences USA_, vol. 84, pp. 8429-8433.

He explains how a true "expert" is better than an "oracle". For example,
in a medical diagnosis application, it is better to know that patients
with such-and-such symptoms have a 75% chance of having cancer than to
know that one particular patient had those symptoms and did (or did not)
have cancer.

Of course, a false "expert" could be much worse than raw data...

     Radford Neal