Extensión del algoritmo de codificación-decodificación basado en PWM para redes neuronales de impulsos

Abstract

Spiking Neural Networks (SNN) are the latest generation of neural networks and attempt to mimic human brain functioning more closely by encoding the information through spike trains. Since most of the real processes are analog, SNN requires the use of encoding-decoding algorithms. The PWM-based encoding-decoding algorithm is a novel temporal encoding algorithm that surpasses its predecessor algorithms in terms of precision. Despite its many advantages, this algorithm requires two chronological values from the original time series in order to encode a spike. In this sense, it is also interesting to be able to apply this algorithm to other types of application, such as image processing, where it does not exist a chronogical order of the points. Hence, this paper proposes an extension of the PWM-based encoding-decoding algorithm, in which is not necessary to employ two consecutive values in the encoding process, enabling the algorithm to be applied to any type of application. In addition, the new extension reduces the computational cost of encoding and decoding processes by more than 50 %.