An Invited Paper for 1997 Seminar on Neural Network Applications in Electrical Engineering

Simulation and Synthesis of Artificial Neural Networks Using Dataflow Models in Ptolemy

Biao Lu (1), Brian L. Evans (1), and Dejan V. Tosic (2)

(1) Laboratory for Vision Systems, Department of Electrical and Computer Engineering, Engineering Science Building, The University of Texas at Austin, Austin, TX 78712-1084 USA
blu@ece.utexas.edu - bevans@ece.utexas.edu

(2) School of Electrical Engineering, University of Belgrade, Bulevar Revolucije 73, 11000 Belgrade, Yugoslavia
etosicde@ubbg.etf.bg.ac.yu

Paper

Abstract

Artificial neural networks (ANNs) are often mixed with digital signal processing (DSP) to form understanding and interpretation systems. These systems are heterogeneous in that they contain a variety of algorithms which may be synthesized onto a variety of software and hardware technologies. Because ANN and DSP subsystems are generally data-driven, their computation and control structure can be unified under a dataflow model of computation. In this paper, we demonstrate that ANN and DSP subsystems can be modeled using dataflow models that yield static implementations on either sequential or parallel machines. We map Hopfield, backpropagation, and biological networks to Homogeneous Synchronous Dataflow (SDF) models. We combine Boolean Dataflow (BDF) and SDF models to model Cellular Neural Networks (CNNs). By modeling DSP operations in SDF, we are free to mix ANNs and DSP subsystems and still retain efficient simulated and synthesized systems due to the static scheduling. We give several examples of simulating and synthesizing mixed ANN/DSP systems using the Ptolemy software environment.

Last Updated 11/07/04.