A Model for Mixed Category Perception Based on Absolute Feature Values - Neural Networks, 1995. Proceedings., IEEE International Conference on

Mixed category perception is necessary in different tasks like recognizing multiple objects in a scene at a time or perceiving music coming from more than one source or even in prediction of disorders from the symptoms of a patient. There exist some models for mixed category recognition from various points of view [l, 2 , 3 , 41. These models accept the features only in terms of their presence or absence, i.e., 0 or 1. The problem of mixed category 'perception in terms of the absolute values of the features has not been dealt with. One of the reasons is that it is relatively a difficult task, as compared to binary input, to predict the resulting feature information after they are mixed. For example, let us have two different categories A : [ f ~ = 4.3, fz = 5.2, f3 = 0 , f d = 01 and B : [fi = 1.2,f2 = O , f 3 = 3 . 1 , f d = 6.51, represented in terms of absolute feature values. In that case, if A and B are superposed then it is difficult to predict how the absolute values of the features will change. On the other hand, if A : [fi = l , f 2 = l , f 3 = 0, f.1 = 01 and B : [fi = l , f 2 = O , f 3 = l , fd = 11 then the presence of absence of the features in the superposed category can be determined by OR logic. Recently, X-tron [5, 61 was developed by considering the task of mixed category perception as a set covering problem where a hypothesis is formed about the presence of a set of objects ivhich would be able to interpret the presence of input features. Instead of presence or absence of a feature, X-tron accepts the degree of presence of the features (i.e., f E [0,1]). In order to represent the absolute feature information in terms of degrees of presence, one neells to have a suitable transformation depending on the ranges of different feature values. Since the ranges of feature values are not knoivn a priori during the on-line categorization process, it is difficult t o compute the degrees of presence, and hence to use X-tron directly with absolute feature values. Here we present a new version of X-tron which will be able to accept the absolute values of the features arid interpret them even in mixed form. The range of absolute information of each ftature is viewed here as consisting of an u11known number of quantized slots. Degree of presence of a feature corresponding to a category is determined with a membership function. An initial guess is made about the size of the slots. Then the network automatically learns the number of slots and the membership function during the