NetNews Usenet Archive 1992 #16/NN_1992_16.iso/spool/comp/ai/neuraln/2988

Text File | 1992-07-27 | 3.3 KB | 75 lines

Newsgroups: comp.ai.neural-nets
Path: sparky!uunet!gatech!destroyer!ubc-cs!unixg.ubc.ca!kakwa.ucs.ualberta.ca!alberta!arms
From: [e-mail] (Bill Armstrong)
Subject: Re: Correctness of NNs
Message-ID: <arms.712266454@spedden>
Sender: [e-mail] (News Administrator)
Nntp-Posting-Host: spedden.cs.ualberta.ca
Organization: University of Alberta, Edmonton, Canada
References: <[e-mail]> <arms.712022732@spedden> <[e-mail]>
Date: Mon, 27 Jul 1992 19:47:34 GMT
Lines: 62

[e-mail] (Robert B Lambert) writes:


>     The individual neurons
>in the brain are causal devices. Their state is based on the history of firing
>of the cell and the history of received pulses.

I would like to agree with you.  If this is true, then you could use
an ALN model to try to learn to approximate this functionality
(boolean signals, discrete time), independently of any non-boolean
parts in the actual system, which merely serve to transform a history
of input pulses to the output pulse stream.  One of my questions is
just: does anyone see a problem in using ALNs this way, even if they
are not an accurate model of the neuron?

>>I will leave it up to the BP people to worry about the safety of their
>>systems.  Up to now, it seems they won't even admit there is a
>>problem.
>>

>I agree with you to a point. Simple NNs by there very nature are unreliable.

That's not my argument.  If we are talking about digital as opposed to
analog NNs, then they can be as reliable as any other computing
device, whether they are ALNs or the usual MLPs.  The problem I see is
that we train them on limited data, and testing will not show up all
the deviations from what we want, even if we can specify what we want.
It's the people who count on miracles that are the real danger in NN
technology.

>What is the future of ANNs? If they are to be used in any situation where a
>fully definable input-output set exists, they have no future as this is the
>application where conventional computer technology excels.

ALNS produce conventional computing technology! Namely combinational
switching circuits, which you could implement using today's cheap,
off-the shelf programmable logic devices. We could use ALN technology
to implement fully definable functions in a faster way.  I see that as
promising for numerical computation of the usual kind.  Here, what
applies to NNs does not apply to BP type nets, because the latter are
not fast enough to compete with conventional methods, while ALNs are.

 If ANNs are to be
>used for real world control and recognition tasks we must face up to the fact
>that such networks while able to give the best performance can never be 100%
>reliable as it is simply not possible to account for all possible inputs.

You have to have a spec, and you have to have a way for proving your
ANN is within spec.  So even in cases where you have astronomical
numbers of inputs, you can still prove an ANN is within spec.  The
major problem is when you have no spec, as in many recognition tasks.

Thanks for your thoughtful comments.

Bill


--
***************************************************
Prof. William W. Armstrong, Computing Science Dept.
University of Alberta; Edmonton, Alberta, Canada T6G 2H1
[e-mail] Tel[phone] FAX [phone]