Fuzzy Vector Quantization for Classification of Olfactory Stimuli from the Acquired Brain Signals

Abstract

Quantization offers a simple means to reduce unwanted data variations into quantized levels of fixed amplitudes. A simple Normalization of data in the scale [0, 1] without attributing special importance to specific sub-intervals in [0, 1] results in significant loss in information, which cannot be retrieved otherwise to identify the proposer class of a given data point. This paper attempts to represent $\boldsymbol{n}$ -dimensional data point patterns into quantized fuzzy vectors of $3n$ dimensions. As the class information for all the data points of the training instances are known, a majority voting over (the individual quantized fuzzy) components of the data points in a class effectively returns an approximate centroidal measure of the data points lying in the same class. The concatenation of such majority-voted components of the data points in a class represents the class centroid of the respective class. The class centroids thus obtained are saved to determine the class of an unknown data point. The classification of an unknown data point here is performed in two steps. First the unknown data point is quantized in fuzzy space similarly as for the training data, and the resulting quantized fuzzy vector is projected along all the unit class centroid vectors of all known classes. In case the projection along the $k$ -th unit class centroid vector is the largest, then the class of the unknown data point is regarded as $\boldsymbol{k}$ . The proposed principle has successfully been applied in classification of olfactory stimuli from the acquired EEG signals of the subjects and the classification performance is found to be superior with respect to the existing state-of-the art algorithms.