Path: utzoo!utgpu!news-server.csri.toronto.edu!rpi!uwm.edu!bionet!agate!ucbvax!ucsd!sdcc6!beowulf!demers From: demers@beowulf.ucsd.edu (David Demers) Newsgroups: comp.ai.neural-nets Subject: Re: Help with network training. Message-ID: <19580@sdcc6.ucsd.edu> Date: 20 May 91 16:53:31 GMT References: <3388@mtecv2.mty.itesm.mx> Sender: news@sdcc6.ucsd.edu Organization: CSE Dept., UC San Diego Lines: 55 In article arms@cs.UAlberta.CA (Bill Armstrong) writes: >barronca@mtecv2.mty.itesm.mx (Ing. Elvia Patricia Barron Cano) writes: >>I have problem with training of the network. I use backpropagation. >>The network have to learn to classify 529 patterns in 3 categories. >>The input is binary, there are 36 inputs, but only two are enable. >Sorry, I don't understand the last bit, if you really mean only two >input leads are enabled of the 36, meaning you only have 4 different >input patterns after the 529 are projected. I shall assume you don't >mean that, but rather that your inputs are restricted to 0 and 1. I too was confused, but upon second reading, a reasonable interpretation might be that only two of the 36 input units are non-zero, thus 1260 (36 choose 2) total possible input patterns. >>First, I test whit 50 hidden units, but the total sum of squares was >>near 10, and to fail in five patterns. I test some times, after with >>more units (52, 55, 60, 80), and two hidden levels, but the results >>were aproximate the same. When I made I tested some rate learning and >>momentums. >>I studied the patterns and I made a network with 19 hidden units. I >>tested with them and 37 units but the response was bad. >>Why it can't find the weights? [...] >Since you have binary input data, presumably what you need to compute >are two boolean functions of 36 variables which separate the space >{0,1}^^36 into three classes. May I suggest you try adaptive logic >networks (ALNs). might be a better tool for this problem. But bp should be able to solve it if there is structure in the classifications. Try reducing the learning rate and boosting the momentum...? I'd suggest using one hidden layer but including direct connections from input to output to easily capture linearities. without knowing more about your data, I can only encourage you to use all tools available. Dave -- Dave DeMers demers@cs.ucsd.edu Computer Science & Engineering C-014 demers%cs@ucsd.bitnet UC San Diego ...!ucsd!cs!demers La Jolla, CA 92093-0114 (619) 534-8187,-0688 ddemers@UCSD