The classification problem in the olfactory system of insects

Ramon Huerta, Thomas Nowotny, Marta Garcia-Sanchez, Henry Abarbanel, and Misha Rabinovich


The olfactory system of the locust has a puzzling variation of coding schemes throughout its various information relay stations. The identity-coded information of the Antenna is compressed into spatio-temporal patterns in the antennal lobe that are relayed towards the mushroom body where they are transformed into a series of quasi-static snapshots of sparse activity. The understanding of these coding schemes is recently starting to emerge. The spatio-temporal coding in the Antennal Lobe seems to be important for at least two important functions: a) learning and memory and b) robust discrimintation of similar odors. In this talk we concentrate on the later stages of processing in particular on the advantages of sparse coding and on an adequate network design for classification purposes. To this end we will for now only consider single snapshots of Kenyon cell activity. The classification task is then accomplished in two steps. The first step is a nonlinear expansion from the Antennal Lobe to the Kenyon cells in the Mushroom Body. This nonlinear projection into a higher dimensional space allows information conservation via sparse coding even if carried out with a non-specific connectivity matrix. In the second phase the sparse code on the Kenyon cell screen is linearly classified in the next processing layer. The neurons that perform this linear classification are hyperplanes whose connections are tuned by local Hebbian learning and competition due to mutual inhibition. We show that the number of Kenyon cells and the degree of sparseness in coding are crucial elements for efficient classification.