Path: utzoo!utgpu!news-server.csri.toronto.edu!rutgers!cs.utexas.edu!sdd.hp.com!elroy.jpl.nasa.gov!decwrl!asylum!osc!jgk
From: jgk@osc.COM (Joe Keane)
Newsgroups: comp.ai.neural-nets
Subject: Re: backpropagation
Summary: Need better initialization?
Message-ID: <4040@osc.COM>
Date: 28 Nov 90 22:48:21 GMT
References: <4916@trantor.harris-atd.com>
Reply-To: jgk@osc.COM (Joe Keane)
Distribution: usa
Organization: Versant Object Technology, Menlo Park, CA
Lines: 14

In article <4916@trantor.harris-atd.com> mlaprade@x102a.ess.harris.com
(laprade maria 42641) writes:
>In my neural nets class we were assigned to build a BPP net to approximate
>the function z = sin(2PIx)sin(2PIy) for
>0<=x<=1 and 0<=y<=1.

This is a fairly difficult function, basically a sinusoidal version of XOR.
It can also be expressed as z = 1/2*(cos(2*pi*(x-y))-cos(2*pi*(x+y))).
Hopefully the nets will learn to use the sum and difference.

>Weights were to be initialized to random values between +-0.1.

I'd say	these weights are too small.  You need to give the net a large amount
of asymmetry to start with, or it will tend to converge to zero.