Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!bb3.andrew.cmu.edu!nntp.sei.cmu.edu!news.psc.edu!hudson.lm.com!news.math.psu.edu!news.cac.psu.edu!newsserver.jvnc.net!newsserver2.jvnc.net!howland.reston.ans.net!tank.news.pipex.net!pipex!oleane!jussieu.fr!math.ohio-state.edu!cs.utexas.edu!swrinde!sdd.hp.com!hplabs!unix.sri.com!nova.sarnoff.com!usenet
From: cds@nova.sarnoff.com (Clay Spence x3039)
Subject: Re: on the relationship between bayesian error bars and the input data density
Message-ID: <DCoy5B.4Dz@nova.sarnoff.com>
Sender: usenet@nova.sarnoff.com (Netnews)
Reply-To: cds@nova.sarnoff.com
Organization: David Sarnoff Research Center
References: <3vlhmj$4bg@yama.mcc.ac.uk>
Date: Wed, 2 Aug 1995 16:07:59 GMT
Lines: 19

In article <3vlhmj$4bg@yama.mcc.ac.uk>, Martin Barry <baz> writes:
> gT(x).Ainv.g(x)
> 
> i was trying to figure out exactly what g(x) was, and how to get it; the paper
> that it is the vector of the derivatives of the output with respect to the
> weight parameters in the network.

I haven't seen the paper. If g is what you say it is, you apply
the usual backpropagation formulas, but at the output units you
set the deltas to the derivative of the squashing function with
respect to the activation of the unit.  (Sorry if there are several
conventions for what delta means.  Precisely, you compute the
derivatives as usual, but pretend the error function is equal to
the output of the net.)

Clay Spence
cspence@sarnoff.com


