Report Number: CS-TR-73-364
Institution: Stanford University, Department of Computer Science
Title: Estimation of probability density using signature tables for appplications to pattern recognition.
Author: Thosar, Ravindra B.
Date: May 1973
Abstract: Signature table training method consists of cumulative evaluation of a function (such as a probability density) at pre-assigned co-ordinate values of input parameters to the table. The training is conditional: based on a binary valued "learning" input to a table which is compared to the label attached to each training sample. Interpretation of an unknown sample vector is then equivalent of a table look-up, i.e. extraction of the function value stored at the proper co-ordinates. Such a technique is very useful when a large number of samples must be interpreted as in the case of speech recognition and the time required for the training as well as for the recognition is at a premium. However, this method is limited by prohibitive storage requirements, even for a moderate number of parameters, when their relative independence cannot be assumed. This report investigates the conditions under which the higher dimensional probability density function can be decomposed so that the density estimate is obtained by a hierarchy of signature tables with consequent reduction in the storage requirement. Practical utility of the theoretical results obtained in the report is demonstrated by a vowel recognition experiment.