Newsgroups: comp.robotics
Path: brunix!cat.cis.Brown.EDU!agate!howland.reston.ans.net!EU.net!uknet!festival!ainews!ruby!simonpe
From: simonpe@ruby.ed.ac.uk (Simon Perkins)
Subject: Re: 100 Billion Neurons Nonsense
Message-ID: <Cs9w35.8GL@aisb.ed.ac.uk>
Sender: news@aisb.ed.ac.uk (Network News Administrator)
Reply-To: simonpe@ruby.ed.ac.uk (Simon Perkins)
Organization: Dept of AI, University of Edinburgh, Scotland
References: <eZ1Noc4w165w@sfrsa.com> <2utltf$2fi@unix1.cc.uop.edu> <2uupb0$kc7@orion.cc.andrews.edu>
Date: Fri, 1 Jul 1994 17:47:28 GMT
Lines: 40

I think there is a real trap in the numbers game approach to comparing
the processing power of brains and computers, namely: What is an
equivalent lump of processing power in each system? So we have, say
10^7 (?) transistors on a Pentium, and say 10^11 neurons in a human
brain. The transistors have switching times of the order of 10^-9
seconds, and neurons `switch' in say 10^-3 to 10^-2 seconds. So
comparing apples with oranges we naively say that the brain isn't that
much more powerful than modern computers after all.

But of course even a single neuron can actually carry out very
specialized and complicated processing so how many transistors equals
a neuron? And then there's connectivity - some neurons are connected
to 10^5 other neurons, and even an average figure may be something
like 10^4. I've heard figures of around 10^15 connections in the
brain. What factor does this massive connectivity and parallelism
introduce - it almost certainly more than compensates (many times
over) for the relative slowness of the individual neurons. So the
naive comparison looks even more shaky than before.

Then we have the bee brain with maybe only a few tens of thousands of
neurons - can we at least beat that? It's probably not as easy as that
number makes out. The point is that the computer's basic units are
very general and abstract things compared to the specialized elements
of the bee's sensory-motor system, and hence you need far more of them
to do the same thing. The physical structure of the bee is responsible
for much of it's intelligence. There are lots of examples of
mechanical processes which perform highly specialized calculations and
processes much more efficiently than an equivalent implementation on a
general purpose computer. As a quick example, consider filling a 2D
flat tray of arbitrary shape by first performing a flood fill
algorithm on a computer simulation of that tray, and then just
sloshing a jug of water into it. Using the physical properties of the
environment can save an awful lot of computation.

-- 
Simon Perkins                             simonpe@aisb.ed.ac.uk

Dept. of AI,
Edinburgh University.
--------------------------------------------------------------------
