Influence of bio-inspired activity regulation through neural thresholds learning in the performance of neural networks

Abstract

Studies on Kenyon Cells (KCs) in the olfactory system of locust have found evidence that some of them are more responsive towards stimuli than others. This variability suggests that there could be a heterogeneous neural threshold distribution among KCs. This work explores the hypothesis that the heterogeneity in neural thresholds could control the activity level of populations of neurons and facilitate the generation of sparse code to achieve a better representation of stimuli. In order to study this hypothesis, an artificial neural network that adapts many of the strategies observed in the locust olfactory system is proposed, including a new learning algorithm capable of finding the best neural threshold distribution to resolve a certain classification problem and a gain control term that keeps the activity level of the neurons in the hidden layer under a certain limit. The bio-inspired neural network gets the best results when the activity level in the hidden layer is low, for which neural thresholds are more important for a better performance than the connections between the neurons in different layers. Also, it was found that the separability of the internal representations of the different patterns in the classification problem improves when the thresholds are adjusted by the learning algorithm in order to control the activity level. Finally, the performance of this network was compared to SVMs and MLP when applied to resolve a complex classification task with data collected by electronic noses and affected by sensor drift, for which the bio-inspired network obtains very satisfactory results, reaching a performance similar or even better than SVMs and MLP.