From: Dale Johnstone (DaleJohnstone@email.com)
Date: Tue Apr 03 2001 - 01:13:30 MDT
Eliezer S. Yudkowsky wrote:
> Anders Sandberg wrote:
> >
> > Well, remember that timing of neural processes is limited by the speed
> > of neural interactions (roughly ~1 millisecond). It is doubtful that
> > precision greater than that is possible without employing many neurons
> > doing clever things.
>
> If I recall correctly, bats can do ~1 microsecond discrimination as part
> of how they process echoes - specifically, IIRC, bats can discriminate a
> one-microsecond difference in echo arrival time between one ear and the
> other. This would, however, be in the category of "many neurons doing
> clever things".
Yes, barn owls also have a temporal precision in the microsecond range
(~5us) which allows them to pinpoint sounds to within 1 or 2 degrees
horizonally. This is pretty impressive considering the spikes in the owls
auditory system has to cope with a temporal jitter of about 40us off their
prefered phase angles.
There's evidence(*) to suggest the barn owl uses a neat Hebbian learning
trick to strengthen those signals in phase, and weaken those that aren't -
and of course tons of neurons to average together.
What I really like about this is that it underlines how hard the brain has
to work just to get a little accuracy over all that noise. An awful lot of
those neurons are just voices in the crowd in various population encoding
schemes. My guess is you could do the same as what the brain currently does
with 2 or 3 magnitudes fewer neurons if they were rate based and reliable.
That's before you even begin to consider the huge speed advantages
artificial neurons have.
Evolution has done an amazing job to get so far. I mean who would have
thought it - computation with living cells? Boy are *we* history!
Regards,
Dale Johnstone.
(*) Discussed in section 14.3 of Pulsed Neural Networks by Maas and Bishop.
http://www.amazon.com/exec/obidos/ASIN/0262632217/qid=986281259/sr=1-1/compu
tronium-20
(Note: This is an affiliate link.)
This archive was generated by hypermail 2.1.5 : Sat Nov 02 2002 - 08:06:48 MST