Research On BP Neural Network Classification Model Of Parallel Multi-mode

Back-propagation neural network which is powerful in self-organization andself-learning can effectively extract latent characteristic and law from the input data ofunknown mapping function, and it has played an important role in the field of patternrecognition, data classification, forecasting. However, in practical application, theclassification performance and time efficiency of single BP neural network is not sosatisfactory when dealing with complex multi-pattern classification problem of large amountof data, and the traditional improved BP algorithm is also difficult to meet the requirement.In the first part, according to the study of BP model structure and analyzing theclassification principles and limitations of single multi-output BP, this paper has proposed amodular MBP multi-mode classification model based on the structure parallel, which hassimplified a multi-mode classification problem into a single-mode classification problem withmultiple modules concurrent execution, thus to improve the training accuracy. What’s more,from the perspective of parallel execution, this paper has designed a structure parallelalgorithm of coarse-grained for the MBP model, based on explicit data and task allocationunder the environment of MPI cluster. The algorithm maps each sub-module unit to differentcluster node, in order to implement MBP sub-modules parallel processing, accelerate thespeed of MBP network and reduce its training time. The experimental results show that theparallel algorithm has achieved an overall average of85.9%recall rate and the accuracy rateof85.3%in the multi-mode image classification, and presents an approaching linear speed-upratio, which proved that the parallel algorithm is correct and effective.In the second part, for large-scale sample data grouping strategy, this paper has proposeda SOM-MBP combination classification model which is based on data parallel. The modeluses SOM network to provide high compactness and regularity sample set for MBP to furtherimprove the classification accuracy. And then according to analyze the SOM-MBP modelcomputational logic and consider the cluster fault-tolerant, the paper has designed the insideparallel neural network with Map and Reduce methods under the Hadoop distributed clusterin the way of migrating from calculation to data, therefore achieving SOM network and MBPsub-module internal parallel processing. On the other hand, the paper has achievedparallelization between MBP sub modules by using parallel multi-job submission. Theexperimental results show that the algorithm has better classification performance, highspeed-up and fault tolerance when processing multi-mode classification of large-scale images,which illustrates the effectiveness and feasibility of the parallel algorithm and reflects that thealgorithm is superior in processing data-intensive problems.