Learning methods for odor recognition modeling

Abstract

This paper presents a phenomenon modeling of human odor perception. For a recognition test by discrimination, odors are presented in couples. Pair odors are separated by an arbitrary defined time interval. Individuals have to recognize if the odor which is presented the second time is the same as, or different to the odor which was smelt the first time. Recognition does not imply identification but only remains validated by success or check terms. For each tested odor, an evocation list is established by various individuals, hence a feature profil for each odor is available. From these data, the aim is to design a prediction model which determines the score for an individual uniquely from his odor evocation pattern. For modeling this phenomenon, we have applied learning methods especially neural networks. A process named “spy” was used to find the best architecture and the best parameters of a given neural network. This tool observes a net of neural networks, which all work in parallel, from the same sets of patterns, but which differ from one another in some of their parameters. In this communication, we also present a comparative study between neural network performances and those given by discriminant analysis which was applied to this odor memorizing problem.