We here present and compare different massively paraIIe1 Implementatlous of multilayer feedfomard neural networks on a MasPar MP-1216, a parallel SIMD computer with 16,384 p r 0 c e . e " For mukllayer feedforward networks we have obtained sustphed rates of up to 348 M CPS and 129 M CUPS with backpropagation, a hlgh mark for general purpose SIMD computers. This paper focuses on the problems of mapping neural networks to parallel hardware, on implementation problems in obtainlng high propagatlon rates on a SIMD machine and on problems with the resulting learning algorithms.Keywords: artlllclal neural networks, neural network simulators, massive parallelism I. INTRODUCTlON AND MOTIVATION Our research group wants to understand the advantages and the trade-offs of the various artificial neural network paradigms and learning algorithms, their training efficiency and generalization capabilities and their suitability for massively parallel implementation. We have developed a neural network simulator, SNNS, which has proven well suited for research on learning algorithms, on issues of visualization, training and performance and on parallel implementation of neural networks. SNNS is also used in a number of other university research groups and with growing acceptance in industry as a neural network evaluation and prototyping tool. In this paper we are describing the experiences we gained in developing a massively parallel simulator kernel for SNNS running on our 16 K processor MasPar MP-1216.
STUTTGART NEURAL NETWORK S W A T O R SNNS (Stuttgart Neural NetworkSimulator) [ll, 121, [31, is an efficient and portable neural network simulation environment for Unix workstations developed at the University of Stuttgart. It is a software tool to generate, train, test and visualize artificial neural networks. The whole network simulator has been developed in C on Unix workstations. The graphical user interface was implemented under X-Windows XllR5 (Athena widget set), for maximal portability.SNNS now consists of a sequential and a parallel simulator kemel and a graphical user interface (Fig. 1). The simulator kemel operates on the internal representation of the neural networks and performs all operations of the learning and recall phase. It is coupled with the graphical user interface via an interface of function calls. The simulator kernel has already been ported to a number of architectures (Sun, HP, DEC, IBM, etc.).
A. Graphical User Inregace of SNNSThe graphical user interface is used to create, visualize and modify the network topology interactively. All display elements are kept in separate windows and thus can be arbitrarily arranged. The user has a powerful set of operations (insertion, deletion, copying, moving) which may be applied to individual units or to selections of units and may affect links as well, like 'copy all selected units with their input links' or 'delete all links into the selected units'. Networks can be modified through the user interface during simulation. Units can be introduced, removed, or have their...