Kang, Miew How
(2016)
Study Of Modified Training Algorithm For Optimized Convergence Speed Of Neural Network.
Masters thesis, Universiti Sains Malaysia.
Abstract
Artificial neural network (ANN) are widely used as an engineering approach to mimic the human brain activities. It has applied in different aspects such as pattern recognition, alphabet or digit classification, handwriting recognition, speech recognition, fingerprint identification, data mining, robots and etc. Backpropagation is the most common artificial neural training algorithm, however it is suffering with the slow convergence rate issue. A study to improve the slow convergence rate of a neural network without sacrificing the accuracy of the network are carried out. In this research, a handwritten character recognition model are implemented in C++ programming with ability to classify digits 0, 1, 2, and 3. This model are built up with 64404 neural network where input data are 8 x 8 dimension image and output are classified to 4 digits which are 0, 1, 2 and 3. Two modified algorithms are proposed in this research, which are mixture of the momentum algorithm with different learning rate algorithms. First proposed algorithm is the combination of momentum algorithm with adaptive learning rate (ALR) algorithm, and second proposed algorithm is the combination of momentum algorithm with automatic learning rate selection (ALRS) algorithm. Convergence rate are showed 18% improvement in the algorithm mixture with ALR algorithm, however there is no significant improvement of timing for algorithm mixture with ALRS algorithm.
Actions (login required)

View Item 