Modeling Receptor Neuron Convergence

To form a chemotopic mapping, we must first define a selectivity measure upon which IR absorption features can be clustered together. In this work, this is accomplished by treating the IR absorption at a particular wavelength across a set of odorants as an affinity vector  

The convergence model operates as follows. The SOM is presented with a population of IR absorption features (corresponding to each wavenumber), each represented by a vector in C-dimensional affinity space, and trained to model this distribution. Once the SOM is trained, each IR absorption feature is then assigned to the closest SOM node in affinity space, thereby forming a convergence map from which the response of each SOM node is computed as  

To help visualize this model. Fig. 6.2. illustrates a problem with absorption spectra of three odors (labeled as A, B and C). The affinities at different wavenumbers are shown 

This convergence model works well when the different sensors are reasonably uncorrelated, since the projection of sensor features across the SOM lattice approximates a uniform distribution, i.e., maximum entropy (Lancet et al. 1993 Laaksonen et al. 2003). Unfortunately, the population of sensors created through IR absorption spectra tends to be over-sampled. As a result, a few SOM nodes tend to receive the majority of input, which capture the common-mode response of the sensor, overshadowing the most discriminatory information. To avoid this issue, the activity of each SOM node is normalized by the number of sensor features that converge onto it  

Note that this solution is not driven by biological plausibility but largely by the limitations of the sensors. 

---

Relating Sensor Responses of Odorants to Their Organoleptic Properties by Means of a Biologically-Inspired Model of Receptor Neuron Convergence onto Olfactory Bulb... 

---