Computing with Arrays of Coupled Oscillators

Abstract

Over the past two years, several experimental results have shed new light on the question of cortical oscillations and their possible roles in neural information processing. In the labeling hypothesis, temporal characteristics such as the phases and/or frequencies of pools of oscillating neurons are used to transiently encode information, in particular to label various features of an object by synchronous activity of the corresponding feature extracting neurons. We propose that two dimensional arrays of weakly and locally coupled oscillators be considered as a new class of architectures, particularly well suited for low level sensory processing such as early vision. Using the coupled limit cycle approach, we investigate several computational properties of such architectures, how they intrinsically differ, for instance, from discrete spin systems, and the importance of two dimensional effects. We show that such arrays are characterized by a very enhanced sensitivity to external inputs and the ability to organize themselves into flexible “patchy” structures over extremely brief transients. Relevance for the neurophysiological data is discussed. Possible applications to early vision, motion analysis, figure ground separation, pattern recognition are demonstrated by coupling oscillator architectures with arrays of filters or other networks.