kirkaas@oahu.cs.ucla.edu (paul kirkaas) (11/29/90)
I am looking for some different types of input/transfer functions on neural inputs. Take a single neuron with a one dimensional input field. I want to it classify two types of very simple input patterns --- those with two peaks, and those with one. It has plenty of inputs; starting with X1 at the left of the input field proceeding to XN at the far right. Each input is either 1 or 0 depending on whether the signal is high or low. I'm not concerned with learning; I will worry about that later. Examples: Two Peaks ________________ __________________ ________| |______| |__ One Peak __________________________________________ ________| |__ Two Peaks __ __ ___________| |_| |__________________________________ One Peak _______ ___________| |__________________________________ Two Peaks __ __ ________________________________________| |_| |______ .... Now, it seems clear that the standard input function, the linear weighted sum of inputs, SUM( X * W), will be unable to divide the two input categories of 1 or 2 peaks. What I am interested in is some other smiple manipulation of input values that could make that distinction --- in other words, some way to do scale and translation invarience in a simple neuron with a simple input pattern. Any suggestions or references? Thanks. Paul -- Paul Kirkaas kirkaas@cs.ucla.edu
landman@hanami.Eng.Sun.COM (Howard A. Landman) (12/07/90)
In article <1990Nov29.023654.9491@cs.ucla.edu> kirkaas@oahu.cs.ucla.edu (paul kirkaas) writes: >Take a single neuron with a one dimensional input field. >I want to it classify two types of very simple input >patterns --- those with two peaks, and those with one. Easy. First preprocess the field by XORing adjacent bits, so you get a spike for each transition: >Two Peaks > ________________ __________________ >________| |______| |__ Becomes: ________||_______________||_____||_________________||_ Then just use the neuron to count spikes and compare against a threshold. Two peaks becomes four spikes, one peak becomes 2 spikes, so a threshold of 3 should give excellent separation. You could also choose to only count rising (or falling) edges. >Any suggestions or references? A very similar problem has often been used as an example in logic synthesis, and is discussed in my paper "Logic Synthesis at Sun" in CompCon Spring 89, p.469-472. Other references are R. Spicklemier (ed.), Oct Tools Distribution 2.1, synthesis examples section, Electronics Research Laboratory, U.C. Berkeley, March 25, 1988; and Synopsys, Inc., Design Compiler Examples, June 28, 1988. -- Howard A. Landman landman@eng.sun.com -or- sun!landman