Computational acceleration using neural networks

The author's recent participation in the Small Business Innovative Research (SBIR) program has resulted in the development of a patent pending technology that enables the construction of very large and fast artificial neural networks. Through the use of UNICON's CogniMax™ pattern recognition technology we believe that systems can be constructed that exploit the power of "exhaustive learning" for the benefit of certain types of complex and slow computational problems. This paper presents a theoretical study that describes one potentially beneficial application of exhaustive learning. It describes how a very large and fast Radial Basis Function (RBF) artificial Neural Network (NN) can be used to implement a useful computational system. Viewed another way, it presents an unusual method of transforming a complex, always-precise, and slow computational problem into a fuzzy pattern recognition problem where other methods are available to effectively improve computational performance. The method described recognizes that the need for computational precision in a problem domain sometimes varies throughout the domain's Feature Space (FS) and high precision may only be needed in limited areas. These observations can then be exploited to the benefit of overall computational performance. Addressing computational reliability, we describe how existing always-precise computational methods can be used to reliably train the NN to perform the computational interpolation function. The author recognizes that the method described is not applicable to every situation, but over the last 8 months we have been surprised at how often this method can be applied to enable interesting and effective solutions.