A Localist Paradigm for Big Data

Big data problems involve more than being able to create a network that can recognize based on a big data set. Big data problems also involve being able to incorporate new information as it arrives. Rehearsing big data sets may require an inordinate amount of time. We present a localist neural network recognition method that can perform equivalent recognition to popular distributed neural networks (shown mathematically) but does not require rehearsing for learning or update. It can also be placed in deep network configuration. However, the focus of this work is not the details of deep networks. The focus is on how easy it is to create and update individual layers. This is an important bottleneck because ultimately creating and updating layers are a problem whether networks are in a deep configuration or not. We use a small laptop running matlab as a microenvironment to reveal data limits determined by limited memory and processing speed. Within this microenvironment we show our approach accepts the largest datasets. Ultimately we encountered limits of the matlab routines to generate random numbers before the limit of our algorithm.