An introduction to the Neural DF architecture

Abstract

Nowadays, artificial neural network models have been largely simulated on conventional computers, proving their ability to solve a large range of complicated problems. The real potential of these neural models will only be available with the development of highly parallel architectures that are designed to optimize the intensive computational requirements of these neural models. However, there exists strong analogy between neural networks and data flow graphs (mainly control of computing in sense data-driven) data flow architectures represents suitable platform for implementation of neural networks. The proposed data flow architecture described in this paper is composed of a number of processing elements that each can be reconfigured to carry out computations of various neurons at run time.