Olfactory sensor processing in neural networks: lessons from modeling the fruit fly antennal lobe.

The insect olfactory system can be a model for artificial olfactory devices. In particular, Drosophila melanogaster due to its genetic tractability has yielded much information about the design and function of such systems in biology. In this study we investigate possible network topologies to separate representations of odors in the primary olfactory neuropil, the antennal lobe. In particular we compare networks based on stochastic and homogeneous connection weight distributions to connectivities that are based on the input correlations between the glomeruli in the antennal lobe. We show that moderate homogeneous inhibition implements a soft winner-take-all mechanism when paired with realistic input from a large meta-database of odor responses in receptor cells (DoOR database). The sparseness of representations increases with stronger inhibition. Excitation, on the other hand, pushes the representation of odors closer together thus making them harder to distinguish. We further analyze the relationship between different inhibitory network topologies and the properties of the receptor responses to different odors. We show that realistic input from the DoOR database has a relatively high entropy of activation values over all odors and receptors compared to the theoretical maximum. Furthermore, under conditions in which the information in the input is artificially decreased, networks with heterogeneous topologies based on the similarity of glomerular response profiles perform best. These results indicate that in order to arrive at the most beneficial representation for odor discrimination it is important to finely tune the strength of inhibition in combination with taking into account the properties of the available sensors.