Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!rochester!udel!news.sprintlink.net!redstone.interpath.net!sas!mozart.unx.sas.com!saswss
From: saswss@hotellng.unx.sas.com (Warren Sarle)
Subject: Re: Scaling the rows or the columns? With which method?
Originator: saswss@hotellng.unx.sas.com
Sender: news@unx.sas.com (Noter of Newsworthy Events)
Message-ID: <DFHJGo.B5@unx.sas.com>
Date: Mon, 25 Sep 1995 23:48:24 GMT
X-Nntp-Posting-Host: hotellng.unx.sas.com
References: <Pine.SGI.3.91.950923091805.5374D-100000@bgumail.bgu.ac.il> <441pkm$4il@newsbf02.news.aol.com>
Organization: SAS Institute Inc.
Lines: 36


In article <441pkm$4il@newsbf02.news.aol.com>, mglinws@aol.com (MGLinWS) writes:
|> What methods are available for resistant standardization, using the median
|> or another indicator of the central tendancy other than the mean ?
|>
|> What measure is used for the scale: median absolute deviation (MAD) ?

The median and MAD are very robust but statistically inefficient. For
some superior alternatives, see:

   Iglewicz, B. (1983), "Robust scale estimators and confidence
   intervals for location", in Hoaglin, D.C., Mosteller, M. and
   Tukey, J.W., eds., _Understanding Robust and Exploratory Data
   Analysis_, New York: Wiley.

|> This issue seems important as many real world data sets are not normally
|> distributed.  For example, I am trying to use neural data (spike trains
|> from real neurons) to discriminate between a set of behavioral events
|> associated with the neural activity.  I would like to try some nnet
|> methods to compliment what I have done already using CART and discriminant
|> analysis.  The distributions of spike counts for most neurons are highly
|> skewed and have different variances at different levels.  What can I do ?

Standardization, robust or otherwise, won't do you any good. Besides,
weird distributions of the _inputs_ are not necessarily a problem. Given
enough data and enough hidden units, the net will make whatever
adjustments are needed. However, it is possible that a nonlinear
transformation, as Horace suggested, may help under less idyllic
conditions. For a highly skewed distribution, try a square root or
logarithmic transformation.

-- 

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.
