NetNews Usenet Archive 1992 #26/NN_1992_26.iso/spool/comp/ai/neuraln/4172

Text File | 1992-11-04 | 1.1 KB | 25 lines

Newsgroups: comp.ai.neural-nets
Path: sparky!uunet!usc!sol.ctr.columbia.edu!destroyer!gumby!yale!yale.edu!jvnc.net!nuscc!ntuix!ntuvax.ntu.ac.sg!cyxzhou
From: [e-mail]
Subject: Question: Forced learning
Message-ID: <[e-mail]>
Lines: 12
Sender: [e-mail] (USENET News System)
Nntp-Posting-Host: v9001.ntu.ac.sg
Reply-To: comp.ai.neural-nets
Organization: Nanyang Technological University - Singapore
Date: Thu, 5 Nov 1992 07:44:52 GMT

We have used a feed-forward backprop network to simulate ore grade
distribution. The problem is very simple. Input: (x,y) coordinates, output:
Z (grade). While the results were very good, there was a problem of scaling
down of data values which did not have a high frequency of occurence. But
reality, very often these values at certain points (x,y) are know to be correct
values. My question: is there anyway to force the NN to remember these values
as being correct in the training process? I would welcome any comments or
suggestioins. You may send your reply to: [e-mail].

Best regards.

Zhou Yingxin