Clustering based on synchronization of pulse-coupled oscillators

We introduce a new clustering approach based on a model of mutual synchronization of pulse-coupled biological oscillators. The proposed algorithm, called Self-Organization of Oscillators Network (SOON), models a set of feature vectors by a population of coupled integrate-and fire oscillators. As the algorithm evolves, it organizes a population of oscillators (or feature vectors) into a set of stable sub-populations (or clusters). Each oscillator files synchronously with all the others within its group, but the sub-populations themselves fire with a constant phase difference. Our proposed clustering algorithm is computationally efficient and has several advantages over existing clustering techniques. In particular, it does not require the specification of the optimal number of clusters, and it is not sensitive to noise and outliers. Moreover since our approach does not involve the explicit use of an objective function, it call incorporate non-metric and non-differentiable distance measures.