Probabilistic-based neural network implementation

This paper addresses a simple way for neural network hardware implementation based on probabilistic methodologies. We propose a new codification scheme that can be considered as an extension of stochastic computing (unipolar and bipolar codification formats), extending its representation range to any real number by using the ratio between two bipolar coded pulsed signals as codification method. Based on this codification, we propose the implementation of different linear and non-linear stochastic computational elements to be employed in artificial neural networks. Also this paper presents the accuracy associated to the proposed processing. The validation of the presented approach has been done with a sample application, (a spatial pattern classification example). The low cost in terms of hardware of the proposed methodology, along with the complexity of the mathematical expressions that can be implemented allows its use for massive parallel computing.