Efficient implementation of connectionist models on MIMD parallel processors using chordal ring topologies

Summary form only given, as follows. A systematic approach is presented for the implementation of connectionist models on message-passing parallel computers. The work concentrates on multiple-instruction multiple-data (MIMD) architectures with no global or shared memory. Information is transferred between processors by high-speed communication channels. Each processor has its own memory for storing both data and programs. Graph theory is used to find efficient topologies for partitioning a neural network while minimizing communications between processors. An efficient routing algorithm and methodology for problem decomposition are presented where the number of neurons and processors for computation may be specified by the user. Interprocessor communications are automatically established using an algorithm based on chordal ring graphs. Simulations of both backpropagation and Hopfield models have been conducted on an MIMD parallel computer with 32 processors.<>