Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!bb3.andrew.cmu.edu!newsfeed.pitt.edu!gatech!newsfeed.internetmci.com!EU.net!peer-news.britain.eu.net!bcc.ac.uk!link-1.ts.bcc.ac.uk!ubjtp80
From: ubjtp80@ucl.ac.uk (Mr M Peatfield)
Subject: Continuous backprop through time
Summary: Why does error rise after target is reached?
Message-ID: <1996Jan16.100612.83558@ucl.ac.uk>
Date: Tue, 16 Jan 1996 10:06:12 GMT
Organization: University College London
Keywords: continuous backprop time
Lines: 15

I'm mapping a temporal sequence (4 eleven bit vectors) onto a static
representation over 8 time steps using continuous backprop through
time. Often the target vector is reached after the first input vector
is presented. However, subsequent input vector presentations appear to
increase the error rate so that, by the final vector presentation, the
output is only distantly related to the target. Could some kindly soul
let me know whether this behavior is to be expected with this
paradigm?

Regards

Mark.

--
Mark Durrant-Peatfield - UBJTP80@uk.ac.ucl Birkbeck College 
