Multicomputer-based neural networks for imaging in random media

The authors describe a novel technique for imaging in random media using a neural network approach based on a modified backpropagation algorithm. Simulation results indicate that it is possible to produce images of simple structures in 2-D media with a reasonable computation time. The present approach is computation-intensive and for this reason the authors have developed a machine architecture and a machine, Kilonode, which is well suited to this class of computing problems, and which can ultimately be produced at a cost which is suitable for commercial application of the neural network algorithms.<>