From candaten@s1.elec.uq.oz.au Fri Jul 1 14:53:27 EDT 1994
Article: 17515 of comp.ai.neural-nets
Xref: glinda.oz.cs.cmu.edu comp.ai.neural-nets:17515
Path: honeydew.srv.cs.cmu.edu!nntp.club.cc.cmu.edu!newsfeed.pitt.edu!gatech!europa.eng.gtefsd.com!library.ucla.edu!agate!msuinfo!harbinger.cc.monash.edu.au!bunyip.cc.uq.oz.au!s1.elec.uq.oz.au!candaten
From: candaten@s1.elec.uq.oz.au (Adrian Candaten)
Newsgroups: comp.ai.neural-nets
Subject: ANN Simulator Survey - Results
Date: 28 Jun 94 06:28:53 GMT
Organization: University of Queensland
Lines: 132
Message-ID: <candaten.772784933@s1.elec.uq.oz.au>
NNTP-Posting-Host: s1.elec.uq.oz.au
X-Newsreader: NN version 6.5.0 #2 (NOV)
ANN Simulator Survey - Results
-------------------------------------------
Number of Participants so far : 60
Distribution : wide distribution, no local trends
Americas : 24 (40.0%)
> US : 21 (35.0%)
Europe : 29 (48.3%)
> Germany : 14 (23.3%)
Australasia & Asia : 7 (11.7%)
Interest in ANNs : wide
Most Common : image and sound classification and
recognition - approx. 27 %.
Others : control, prediction, learning structures, game
theory, modelling industrial processes, compression
algorithm.
Approach of simulating ANN :
using general-purpose language - 36 (60.0%)
using ANN simulator - 13 (21.7%)
using both (their vote is split) 11 (18.3%)
Reasons stated for using a language :
flexibility, speed, total control of simulation, simpler,
doing new research, helps in understanding the theory,
tight budget and resources, simulators don't see networks
the same way as I do, languages are more available and
portable, most simulators require C code anyway.
Reasons stated for using a simulator :
don't waste time (very popular), doing simple stuff, why
reinvent the wheel, simulator code is better and has less
bugs than mine, can't program.
Most popular languages mentioned:
C - 29/47 (61.2%)
C++ - 16/47 (34.0%)
Others - 12/47 (25.6%)
- pascal(3), fortran(2), objective-C, ML (functional
language), FAST, LISP, C*,..
n.b. people sometimes quota using more than one language. C
was often used along with C++.
Most popular reasons given for using C :
I know it, it is everywhere and cheap, C is fast, but
hard to use, sort of industry standard, good for creating
libraries.
Most popular reason given for using C++ :
I know C and OOPLs suits ANN simulations ie. the nature
of the program reflects the nature of the problem.
Most popular simulators mentioned:
SNNS - 12/24 (50.0%)
NeuralWorks - 3/24 (12.5%)
xerion - 3/24 (12.5%)
NeuDesk - 3/24 (12.5%)
BrainMaker - 2/24 ( 8.3%)
PlaNet - 2/24 ( 8.3%)
Others - 10/24 (41.7%)
- NeuronWindows, Asprin/MIGRAINE, DynaMind, MacBrain,
Nets, Neuralyst, Neuroforecaster, NeuronShell, SESAME.
- Mathematica, MatLab NN Toolbox, Maple
- a few people have written their own language.
n.b. people often use more than one simulator. Because of
their inflexibility, I presume.
Most popular simulation platform :
X -windows - 39/60 (65%)
MSDOS on PC - 17/60 (28%)
UNIX w/o X - 8/60 (13%)
Windows (3.1 or NT)PC - 6/60 (10%)
OpenWindows - 3/60 ( 5%)
Mac - 2/60 ( 3%)
n.b. 21% of researchers use more than one platforms.
Backgrounds :
w/- Engineering (mostly electrical) 27/60 (45%)
w/- Computer Science (CS) 28/60 (47%)
w/o Engin. or CS 8/60 (13%)
n.b. persons may have both -> percent don't add to 100.
VISIBLE TRENDS and POINTS OF INTEREST
-------------------------------------
Between approach and location :
85% (18/24) of Americans(North and South) tend to write
their own simulators, while only 8% (2/24) strictly use a
simulator. This differs from Europe, where 50%(14/28) write
their own and 21%(6/28) strictly use a simulator. In Germany,
the distribution is very even with 36%(5/14) writing their
own, 28%(4/14) using a simulator and 36%(5/14) using both.
Between simulators and location :
SNNS is twice as popular in Germany (67% or 6/9) as it is
outside Europe (33% or 3/9). (p.s. I know it is a small sample
size).
Between languages and platform : none
Between approach and interest :
It seems that sometimes simulators are used to quickly
determine the viability of using a neural network in a
physical system to avoid time-consuming code-writing.
Between approach and background :
If a person has an engineering background, then they will
probably write their own simulation (probability of 70.0% or
19/27). In contrast, a person with computer science and no
engineering background is 61% (16/26) probable of writing
their own. A person with neither of these backgrounds is
likely to use a simulator (3/7) or both (2/7).
-------------------------------------
NOTE :
Sorry for the delay of these results, especially
to those of you that were very prompt with your replies.
I've been getting carried away with my own blah blah research.
As far as my own interest, these results provide enough of
an accurate picture. For a clearer picture, more participants
would be required. I WILL send an update, IFF I receive a sufficient
number of replies.
If you disagree, agree or just are interested with the results,
have your say by filling in the survey forms provided in the news
group and e-mailing me (unless of course, you've done so already).
Thanks
Adrian Candaten (candaten@elec.uq.oz.au)
(update on coffee : number of persons collected - 1)