Efficient Implementation of Correlation Matrix Memories on SN0 based ccNUMA Parallel Computers

Correlation Matrix Memory (CMM) is a single layer binary neural network. One of the principle motivations behind the development of binary neural networks was the ease of implementation. Almost all of todays computing machines operate using binary instructions on binary data and therefore lend themselves well to the simulation of binary networks. However as new, more complex, applications are found for neural networks the size of CMM required the amount of data they operate on can grow very fast. This project looks at methods for fast, efficient and scalable simulation of large CMMs on ccNUMA (cache-coherent non-uniform memory access) machines, and in particular, on on SGI Origin servers. The primary goal being to generate a highly efficient and practical set of algorithms for such simulation. Within this project algorithms for both sequential and concurrent execution are devised and analysed in detail. Different method of concurrency are suggested and results of experiments, carried out on real systems, are presented.