|
The goal of this research is to enhance the capability and to evaluate the performance of the MPANNS (Multiprocessor Artificial Neural Network Simulator). The MPANNS hardware and its basic control firmware was built by a former graduate student in our laboratory. It includes 85 single-chip microprocessors interconnected by tree-like hierarchical shared buses. On the performance evaluation side, we first measure the actual performance data by running real programs on MPANNS. Then we use a software modeling and simulation tool to analyze the system. The results are very useful in showing the directions for future improvements. Due to several constraints confroned by the previous designer, the original MPANNS only simulates Hopfield neural networks. Our first task is to lift this limitation by changing and enhancing its firmware. The result has made MPANNS into a truly general-purpose neural network simulator.
|