The Department of Biophysical and Electronic Engineering (DIBE) at the University of Genova can build on a vast experience in neural network research. Since January 1997, the DIBE team is working on the ESPRIT Programme funded RAIN project for the realisation of four demonstrators of artificial neural network (ANN) applications through the use of High Performance Computing and Networking (HPCN) technologies. For this purpose, they have chosen the popular Multi Layer Perceptron (MLP) network and developed a companion learning algorithm, referred to as Back-Propagation (BP). This cost-effective architecture will operate through a WWW-based interface, any Java-enhanced browser is able to administer, and constitutes an appealing solution for a variety of industrial and medical applications.
In fact, the target architecture for RAIN, which stands for Redundant Array of Inexpensive workstations for Neurocomputing, constitutes a heterogeneous cluster of conventional high performance workstations. The system has to be redundant since the need for vast computational power implies the unavailability of some computing nodes at certain times. These nodes are linked together in an array, as to form one single High Performance Computing node. The use of HPC technology is costly but if you are able to apply it on inexpensive workstations and combine it with the wide-spread ANN techniques, it should be within everybody's financial reach.
Currently, the system consists of seven workstations, four of which are HP VECTRA running on Windows NT, and three of them are "UNIX". The cluster can use three sorts of interconnection networks, namely Ethernet (10 Mbit/s), 100VGAnyLAN (100 Mbit/s) and ATM (155 Mbit/s). There are equally three different software layers: the low level library at the bottom includes a set of optimised neural routines; the communication library provides data exchange between the computing nodes; and the ANN algorithms at the top level rely on both the communication layer and low level library to perform fast core operations. Access to and management of the demonstrators happens through the Java-based user interface.
The architecture allows to fully exploit the system's memory hierarchy because the software tools are installed with a high degree of locality. Thus, the performance of the cluster can be upgraded to the sum of the peak performances of each workstation. One of the machines in the cluster has the role of the Master as it supervises the cluster's functionality. It is also forming a bridge between the Cluster level and the Server level, in order to collect and organise the data which has to be exchanged with the user at the Client level. Here, referral to a remote object is possible as if it were present on the local computer, thanks to the Remote Method Invocation (RMI). While the user is executing distributed neural applications, residing on the cluster, he is simultaneously being informed about the system and its currently running processes.
The RAIN demonstrators are bound to pay excellent services in medical and industrial fields because of the scalability of the workstation cluster, in which new nodes easily can be added to the network. Especially in medical environments, a low cost workstation cluster often forms the only solution to supply the necessary high performance computational power for the ANN applications. Furthermore, the demonstrator allows the user to forecast the power he needs for his particular problem as well as the forthcoming costs. The distributed computational paradigm enhances the reliability of the system, since one failing computing node can be supplied by the others, causing only a graceful degradation of performances without any fatal loss.
The DIBE team has adopted standard programming languages and libraries for the sake of easy transportation between different architectures and computing environments, as to turn RAIN into a portable and flexible system for general purpose functions, such as medical diagnosis, complex processing of images, statistical validation, introduction of clinical cases in a database, high speed character recognition, and so on. You can find more information on HPCN-ANN applications' analysis as well as technical details about the architecture on the Web site of the RAIN project.