AN ANALYSIS OF A CLASS OF PATTERN RECOGNITION NETWORKS 
by LAVEEN KANAL 
General Dynamics/Electronics 
Rochester, New York 
ABSTRACT 
Recent articles show that considerable 
research is currently being devoted to the 
realization of Pattern recognition by a 
class of adaptive networks. This paper 
presents a few results of studies aimed at 
finding the relationships between work in 
this area and various classification 
procedures, and seeing how efforts in 
machine recognition of patterns can be 
reconciled with gome formal principles in 
decision making. 
INTRODUCTION 
Research in classification procedures 
has been largely concerned with two 
situations. One, where an individual 
(pattern, sound) is to be assigned to one 
of k groups, with the numerical value of 
k known but information on the probability 
distributions of observables ranging from 
complete ignorance of the functional form 
of the distribution to the case where the 
functional form and all parameters are known. 
The other situation is similar except that 
the value of k is unknown. The first case is 
typical of many simple pattern and limited 
vocabulary speech recognition situations. 
Thus for example, it may be desired to 
recognize a sound as being one of k known 
sounds, The second situation may be en- 
countered for example, in the identification 
of targets from underwater return patterns 
when only limited information is available. 
Only the first situation is considered here. 
Let a pattern be specified by the results 
of a number of measurements or tests, X59 
i*l,2,...N. In any given instance, the results 
of the tests or measurements may be represented 
by x= (x1, XoyeeeXy)- The case where the x; 
are either O or 1 is considered here. This 
situation would obtain for instance when a 
pattern is placed on an “artificial retina" 
with the outputs of the retina elements being 
quantized in this manner. The cases 
where the x,, iwl, 2,..N can take on a 
number of discrete values and where the xX, 
are continuous are considered eisauienan: 
Let p (xy, XneeeeeXy) denote the joint 
probability distribution of the x; in a given 
group. In section 1 a parametric representation 
for this distribution is presented. In this 
context the classification cabaébilities of 
some proposed pattern recognition networks are 
"Superior numbers refer to similarly numbered 
references at the end of this paper." 
340 
