Abstract
This paper deals with a perceptron-type element in which variable weights change automatically following a certain rule of growth. An analysis of its dynamic behavior is described together with some simulation results.
The element is a summing device. Its output y(t) is a weighted sum of its inputs xi(t), i=1, 2, ……, N, i.e.,
y(t)=NΣi=1wi(t)xi(t)
The inputs xi(t) are assumed to be zero-mean signals, but not restricted to binary signals. Changes in the weights wi(t) are described by the differential equations
Tdwi(t)/dt+wi(t)=axi(t)sgn[y(t)] i=1, 2, ……, N.
A detailed investigation of the solutions of the above equation shows that the element has a strong tendency to separate its inputs into a family of principal components and to pick out the greatest component as its output. This property enables it to perform a variety of types of information processing such as factor analysis, signal filtering, pattern dichotomy, majority decision logic and memory.
