| Artificial Neural Networks (ANNs) represent a new approach to artificial intelligence, but ANN training algorithms create concurrent operation bottlenecks. Some popular conventional training algorithms such as back-propagation have been used in a wide variety of applications, but they are criticized as being biologically implausible, they also suffer from slow convergence and lack of robustness, and they are not suitable for on-line operation.;In this dissertation, a concurrent information processing (CIP) model is provided as an alternative to conventional neurocomputing. Contrasts between the CIP approach and the conventional approach are discussed for a pattern recognition problem in a concurrent learning context. Results indicate that CIP learning per trial can range from 10;The key feature of CIP systems is their ability to learn computing relationships automatically and concurrently. Other useful features include available sequential software that can process incoming records at fairly fast rates; available parallel functions that can process records at very fast rates; provisions for concurrent learning, missing value imputing, and measurement deviance monitoring; and provisions for occasional model refinement and interpretation.;Optimal model refinement is important because it maximizes CIP speed and accuracy when memory specifications are fixed, and it minimizes memory requirements when speed and accuracy specifications are fixed. CIP refinement must be fast as well as optimal, because it must precede optimal CIP operation. Results indicate that optimal models can be quickly identified that produce comparable correct classification rates to full-scale models for a pattern recognition problem, but much more quickly and compactly. CIP fast model refinement procedures include a feature joining method, a feature pruning method, and a feature grafting method. CIP joining methods can quickly cluster features that are redundant, CIP pruning methods can quickly exclude features that are not necessary, and CIP grafting methods can introduce higher-order features that may improve operational accuracy.;The CIP kernel (CIPK) algorithm involves several large matrix and vector operations. A hypercube implementation of the CIPK algorithm is also described in this dissertation. The parallel CIPK algorithm reduces the computing time for these operations and reduces host computer memory requirement, allowing large problems to be treated. Implementation details and a parallel CIPK formulation are given; benchmarks for computing times are provided indicating that the parallel CIPK algorithm is highly efficient; and processing times are obtained to indicate that CIP system performance improved if optimal architecture can be implemented. |
