Extracting functional clusters of glomeruli in rat olfactory bulb by non-negative matrix factorization

Ensemble coding in the early olfactory pathway has been extensively investigated using imaging techniques. These studies have shown that glomeruli with similar affinity gather in close proximity in olfactory bulb, forming a module. In this work, we propose computational methods for analyzing this neural code. Specifically, we show how non-negative matrix factorization (NMF), a machine-learning method for extracting the intrinsic parts of objects, can be used to automatically extract glomerular modules from a database of bulbar activity patterns, as measured with 2-deoxyglucose. The modules extracted by NMF correspond to localized areas in olfactory bulb, in consistency with experimental results from imaging studies on glomerular activity. To validate the emerging representation, we analyzed the relationship between neural activity on these modules and perceptual descriptions of the odorants. We first used pattern-classification techniques to predict ten perceptual descriptors for 53 odorants from their activity on the modules. Our results indicate that NMF is able to extract modules that are intrinsic to the odor coding mechanism. Furthermore, we used mutual information to analyze the relationship between modules and olfactory perception. This analysis revealed the contribution of each module to the olfactory percepts.