Adaptation of Spike-Timing-Dependent Plasticity to Unsupervised Learning for Polychronous Wavefront Computing

Non-Von Neumann computational architectures have lately aroused significant interest as substrates for complex computation. One recent development in this domain is the Polychronous Wavefront Computing (PWC) computational model based on multiple wavefront dynamics. This model is an abstraction and simplification of the artificial neural network paradigm based on temporal and spatial patterns of activity in a pulse propagating media and their interaction with transponders. While this framework is capable of computing basic logical functions and exhibiting interesting dynamic behaviors, methods for unsupervised training of the framework have not been identified. The lack of input weights and the spatio-temporal nature of the PWC framework make direct application of weight adjusting learning methods (e.g., backpropagation) impractical. The paper will describe research into unsupervised learning for PWCs inspired by Spike-Timing-Dependent Plasticity (STDP) methods used with other types of polychronous models. The method is based on adding Leaky Integrate-and-Fire semantics to the PWC framework allowing analysis of activating wavefronts and determination of the optimal location for future stimulation. The transponder's location is then incrementally adjusted to improve its future response. The paper will discuss the learning approach and examine the results of applying the method over a series of stimulations to sample configurations.