Perceptual Grouping through Competition in Coupled Oscillator Networks Martin Meier, Robert Haschke and Helge J. Ritter * Bielefeld University, Neuroinformatics Group 33615 Bielefeld - Germany Abstract. In this paper we present a novel approach to model perceptual grouping based on phase and frequency synchronization in a network of coupled Kuramoto oscillators. Transferring the grouping concept from the Competitive Layer Model (CLM) to a network of Kuramoto oscillators, we preserve the excellent grouping capabilities of the CLM, while dramatically improving the convergence rate, robustness to noise, and computational performance, which is verified in a series of artificial grouping experiments. 1 Introduction The ability to robustly group related perceptual items to form higher-order con- cepts is crucial for many cognitive tasks. Exploiting the recurrent dynamics of neurons, the Competitive Layer Model (CLM) [3] has proven to solve a broad spectrum of complex grouping tasks in a very robust fashion – even in the pres- ence of strong noise. Amongst others, these tasks include segmentation of cell images [4], grouping of object contours in edge images [5], as well as motion segmentation [6]. However, a major drawback of the CLM for real-world appli- cations, is its high demand for computational resources: The network converges slowly and each update step is costly. Furthermore, the network can only hardly escape from a reached optimum, when the grouping dynamics is changed. Hence, inspired by the fast synchronization ability of coupled oscillator net- works [7, 1, 2], we transfer the grouping principles of the CLM to a network of Kuramoto oscillators [7] in order to improve the computational performance. In the new model, each oscillator represents a distinct input feature from an arbitrary feature domain. The coupling strengths between the oscillators are based on the compatibility of the corresponding features. Similar features have a positive compatibility, therefore the corresponding oscillators phase-lock and form a perceptual group, which repels dissimilar features by means of negative couplings. The Kuramoto model has been investigated in many variations [8, 9] and we refer the interested reader to this work. In the following sections, we shortly outline the principles of the CLM and introduce our approach to transfer them to a network of coupled Kuramoto oscillators. In section 4 we evaluate both approaches with regard to the group- ing quality and convergence speed in the presence of increasing levels of noisy connections. Finally, the results are discussed. * This work has been conducted within and funded by the German collaborative research center “SFB 673: Alignment in Communication” granted by DFG.