NetNews Usenet Archive 1992 #16/NN_1992_16.iso/spool/comp/ai/neuraln/2991

Internet Message Format | 1992-07-27 | 3.5 KB

Path: sparky!uunet!van-bc!mdavcr!garry
From: [e-mail] (Gary Holmen)
Newsgroups: comp.ai.neural-nets
Subject: Re: Neural Nets and Brains
Message-ID: <[e-mail]>
Date: 27 Jul 92 23:37:12 GMT
References: <[e-mail]> <arms.711935064@spedden> <[e-mail]>
Organization: MacDonald Dettwiler, 13800 Commerce Parkway, Richmond, BC, Canada  V6V 2J3
Lines: 55

In article <[e-mail]> [e-mail] (Mike Black) writes:
>In article <arms.711935064@spedden> [e-mail] (Bill Armstrong) writes:
>>
>>The theory of multilayer perceptrons as used in BP is horrifying.  You
>>need Kolmogorov's theorem just to show you can do everything you want
>>to approximate continuous functions, but nobody can apply it in
>>practice.  In contrast, every logic designer who has heard of CNF and
>>DNF finds it *obvious* that a adaptive logic net can synthesize any
>>boolean function.  So BP nets lose in a BIG way on the theory side, sorry.
>>
>I couldn't care less about theory...only results...

This seems to be an 'end justifies the means' attitude. How can you assure
your customers that you are producing *SAFE* software without looking at
theory? One of the big concerns I hear consistently from customers is the
neural networks just seem to be 'black boxes'. They want to minimize the risks
they're involved in and a response like " But it worked last time...." just
doesn't do it. I'm not down grading the importance of research into new fields
but maybe we should also be concerned with explaining what we are doing at
the present time. While results are great I would feel much more secure knowing
that there is theory backing up what I'm doing.

>I took the atree software and (for times sake) reduced the multiplication
>problem to the 1 and 2 times tables.  I removed 1*6 from the table and
>let atree crank.  When I tested 1*6 it gave me an answer of ~35.  I do
>NOT call this superior as the backprop net I trained gave me an answer
>that was at least BETWEEN 1*5 and 1*7.   The other problem I ran into
>was running out of memory (16 meg + 64meg swap space) on a problem that
>I had previously solved with backprop.

After reading this message I decided to go home and run the ALN software
on my IBM PC. It is only a 25 Hz 386 with 2 Meg of memory and the
package seemed to work fine. I trained 3 trees on the times tables and
it finished in about 45 mins. The answer I recieved was 1*7 which is about
the same accuracy I've seen from the BP algorithms I've used.

My guess the problem with your answer had to do with the
random_walk portion of the algorithm rather than the ALN's themselves.
I believe that this has been noted for modification in future releases.


In conclusion I would like to note that even though ALN's do have some
disadvantages (not enough of a user base being one of them) that they do
show promise in several areas and I believe that we shouldn't 'throw the
baby out with the bath water' just because it doesn't have the recognition
that BP nets do. Look at them objectively and learn their advantages
and disadvantages so that we can use neural networks to their full potential.

----------------------------------------------------------------------------
- Garry Holmen        |    To Boldly Go Where No Computer Has Gone    -
- Software Engineer     |                  Before               -
- MDA  Richmond, BC     |                           -
----------------------------------------------------------------------------
- email: garry @mda.ca                               -
----------------------------------------------------------------------------