Study of the enhancement of computational capabilities in an olfactory system model by means of neuronal diversity and gain control mechanism

• Autores: Aarón Montero Montero
• Directores de la Tesis: Francisco de Borja Rodríguez Ortiz (dir. tes.)
• Lectura: En la Universidad Autónoma de Madrid ( España ) en 2019
• Idioma: español
• Enlaces
  • Tesis en acceso abierto en: Biblos-e Archivo
• Resumen

  • Every time you inhale, you introduce molecules from the air inside your body. Some of these molecules interact with neuronal receptors located in the olfactory epithelium, which generates a nerve impulse that travels through the olfactory tract to your brain. A direct trip unlike that carried out by the rest of the senses, that have to go through the thalamus in the first place. Hence, odor information can be perceived earlier, alerting us, for example, of potential dangers such as something burning close to us. This fast processing together with the human capability of recognizing a trillion odorants, makes the olfactory system a tool of great interest to study, since it could bring great innovations to the field of pattern recognition and machine learning.

    However, instead of simulating computationally the human olfactory system, we have focused on a simpler and better known system from the point of view of biology, the insect olfactory system. This system allows us to study what properties of its neural network are linked to its ability to classify odorants and extract conclusions that can be extrapolated to our own olfactory system.

    The objective of this thesis is to analyze the role of neuronal heterogeneity in odor discrimination through the computational model of the olfactory system of insects. Specifically, we will analyze the role of three types of neuronal heterogeneity observed in this olfactory system, which are related to each other. In the first place, the existence of varying neural thresholds, which we have demonstrated that allow improving the classification results with respect to the use of the same neural threshold for all neurons. Secondly, specialist and generalist neurons, which respond differently to input stimuli because of their neural threshold values. For this neuronal diversity we observed that although specialist neurons are crucial for the classification of patterns, there is an optimal ratio of specialists/generalists for this classification based on the complexity of the input patterns. Finally, the gain control mechanism of the antennal lobe of insects, produced by the interaction of two populations of inhibitory neurons. We observed that using this inhibitory heterogeneity, to model this gain control mechanism, we are able to simulate more properly the biological behavior of this mechanism and obtain a better classification result with respect to modeling it through a homogeneous inhibition.

    Thus, because whenever we apply the neural heterogeneity in our computational model we obtain better classification results, we suggest that including these mechanisms in artificial neural networks, chemical sensors and other tools for solving pattern recognition problems can be beneficial.