1. Data and variables.

As a data set we used features from 1000 images that have been used for content-based image retrieval in [6]. Each image is represented by a vector of 144 features (variables). The features fall naturally into three categories related to their origin: FFT, RGB, and texture. It is reasonable to assume that dependences between variables in different categories are weak. Thus, the data provides an ideal validation for both the ${\rm IVGA_{VQ}}$ algorithm as well as the general IVGA approach. For the experiments we chose randomly a subset of 50 features: the variable set A consisted of 27 FFT features (numbered 1-27 below), B of 7 RGB features (28-34) and C of 16 texture features (35-50).