Parallel implementation on improved error signal of backpropagation algorithm

The research work presented in this thesis is a continuation of Shamsuddin's work regarding proposed error signal for the backpropagation (BP) algorithm. The main focus is to parallelise Shamsuddin's work in order to improve the speedup of the BP algorithm. The experiments are implemented using the Sequent Symmetry SE30 parallel machine. The BP algorithm uses the data partitioning method with columnwise block striped and the batch mode weight updating strategy. Twenty-six patterns consisting of uppercase letters from 'A' to 'Z' are tested in the experiments. Two main factors taken into consideration in this, experiments are the execution time and speedup and the recognition rates. Shamsuddin's proposed BP parallel version, is compared with the sequential version. Experimental results shows that the execution time of the parallel version is much less than the execution time of the sequential version. The parallel version produces a good speedup as the number of processors, are increased due to the value that is near the ideal value. Experiments for testing the recognition rates involves the twenty-six trained sample data with perfect pattern and untrained sample data with 10% corrupted pattern. The recognition rates results show 100% accuracy for the trained and untrained data using the standard BP and Shamsuddin's proposed BP running sequentially.

Text FSKTM_2001_10 IR.pdf Download (398kB)

Actions (login required)

View Item