Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!bb3.andrew.cmu.edu!newsfeed.pitt.edu!gatech!news.mathworks.com!newsgate.duke.edu!interpath!news.interpath.net!news.interpath.net!sas!newshost.unx.sas.com!saswss
From: saswss@hotellng.unx.sas.com (Warren Sarle)
Subject: Re: Training times. . . again
Originator: saswss@hotellng.unx.sas.com
Sender: news@unx.sas.com (Noter of Newsworthy Events)
Message-ID: <E6upLF.ApF@unx.sas.com>
Date: Mon, 10 Mar 1997 23:36:51 GMT
X-Nntp-Posting-Host: hotellng.unx.sas.com
References:  <mungman-ya02408000R0803972054110001@news.intersurf.com>
Organization: SAS Institute Inc.
Lines: 60


In article <mungman-ya02408000R0803972054110001@news.intersurf.com>, mungman@intersurf.com (Edwin Kirschner) writes:
|> ...
|>    As I understand it the biggest limitation to neural network usage is
|> training time.  (If I am wrong would someone please correct me?)  With that
|> in mind I was wondering a few things:
|> 1. How large are typical data sets that are used in commercial neural
|> networks.  By how large I mean: how many inputs,

For typical business and industrial applications, anywhere from one to
thousands.

|> how many outputs,

Often just one, for OCR dozens. I can't think of any common
application that would have more than a couple hundred outputs (maybe
someone else can).

|> and how
|> many instances of these input/output sets are trained into a given neural
|> network?

Anywhere from a couple dozen to a couple million.

|> 2. For a given data set of a specific size, how long does it typically take
|> commercial software to train that data and make the network ready for data
|> processing?

That depends very much on how complex the data set is and how complex
a network you train.

|> For example it would obviously take less time to train the XOR
|> problem then the sonor data at the CMU benchmark site.    Can anyone tell
|> me how long it takes to train neural networks from this perspective or
|> should I restate the question again.

Typical times would be a couple seconds for XOR, a couple minutes for
the sonar data, and a couple hours for the "adult" census data set in
the UCI collection on my ancient HP workstation.

|>    The motive behind my question is this.  I have developed a neural
|> network that I think is pretty fast  however I don't know how it compares
|> to other neural networks and this seemed like a good place to find out. 
|> Thank you in advance to whoever responds.

The first question is, how well does your network generalize?  There's
not much point in comparing training times without taking into account
generalization accuracy. The more time you spend training, the better
generalization you can get, up to a point that depends on the
complexity of the problem. The methods that provide the best
generalization, such as Bayesian learning via Markov chain Monte
Carlo, usually take the most training time.

-- 

Warren S. Sarle       SAS Institute Inc.   The opinions expressed here
saswss@unx.sas.com    SAS Campus Drive     are mine and not necessarily
(919) 677-8000        Cary, NC 27513, USA  those of SAS Institute.
 *** Do not send me unsolicited commercial or political email! ***

