Path: utzoo!utgpu!news-server.csri.toronto.edu!cs.utexas.edu!uunet!tdatirv!sarima From: sarima@tdatirv.UUCP (Stanley Friesen) Newsgroups: comp.ai.philosophy Subject: Re: Continuous vs discrete Message-ID: <195@tdatirv.UUCP> Date: 16 Apr 91 15:13:21 GMT References: <19175@ogicse.ogi.edu> <19392@ogicse.ogi.edu> <19628@ogicse.ogi.edu> Reply-To: sarima@tdatirv.UUCP (Stanley Friesen) Organization: Teradata Corp., Irvine Lines: 37 In article <19628@ogicse.ogi.edu> maxwebb@moe.cse.ogi.edu (Max G. Webb) writes: >Possibly, but if these systems are so computationally nonrobust >that rounding and discretization errors in a computer simulation >obliterate their value, then how could they have evolved? Keep >in mind that the neurons and the architecture of the nets have >been changing and evolving at the same time? Quite right. I suspect that one of the most important factors controlling neural evolution is stability in the face of noisy input. Ineed in some sense that could be said to be the primary dunction of the nervous system. >Actually, no. The tremendous range of light values detectable >(1 photon up to a sunbeam) is compressed to a narrower range >before the nervous system ever sees it, by the photoreceptor >(first of all). Secondly, there is plenty of evidence that it >is _edge_ information that is passed back, possibly other compressions >of the data. An article in the *latest* issue of Scientific Amerian (April, 1991) is relevant here. The primary encoding coming out of the eye seems to be local relative intensity on a logarithmic scale. (This does indeed amplify edges since they inovle abrupt changes in light levels). That is V = log(X/X-bar). (Where X-bar is the local weighted average of the light intensity). > While i also like floating point, >you would have a very hard time convincing a biologist that >a change of 2^-24 in the operating levels of a neuron would >destabilize the system! In fact one of the features of the eye circuitry is its stability over a wide range of inputs levels, and its ability to extract detail from limited, noisy data. -- --------------- uunet!tdatirv!sarima (Stanley Friesen)