Newsgroups: comp.robotics
Path: brunix!cat.cis.Brown.EDU!agate!ihnp4.ucsd.edu!swrinde!cs.utexas.edu!convex!news.duke.edu!godot.cc.duq.edu!nntp.club.cc.cmu.edu!honeydew.srv.cs.cmu.edu!news
From: hpm@cs.cmu.edu (Hans Moravec)
Subject: Re: 100 Billion Nuerons
Message-ID: <Csw2xD.G9F.3@cs.cmu.edu>
Sender: news@cs.cmu.edu (Usenet News System)
Nntp-Posting-Host: cart.frc.ri.cmu.edu
Reply-To: hpm@cs.cmu.edu
Organization: Field Robotics Center, CMU
References: <ZByDPc2w165w@sfrsa.com>
Date: Wed, 13 Jul 1994 17:22:17 GMT
Lines: 26


Brad Smallridge, bsmall@sfrsa.com writes
re accumulating neural inputs over time:
>
>I'm not sure. When I originally thought of this kind of system
>I assumed I would need a mechanism to bleed off the counters.
>What a pain though because you would need to touch every cell
>in the system periodically. Another way is to somehow time stamp
>the cells counters and when a new stimulus comes into the cell,
>you ask if the old counter value is too old, and make a decision
>based on the count and how old the count is.
> 
>I'm thinking now that you really don't need to bleed off the 
>counters at all. So what if a seldom visited cell ocassionally 
>goes off. In the grand scheme of things, this once in a great 
>while shouldn't effect the outcome of the thinking process. It's
>the cells that are firing a lot that really matter.

Another possibility is to do the "bleed" probabilistically.  Decrement
just a few of the cells, chosen at random, or on some scrambled round
robin schedule.  The slower the bleed rate, the fewer cells you need
to touch each cycle.

			-- Hans Moravec   CMU Robotics


