ADAPTIVE LEARNING RATE VERSUS RESILIENT BACKPROPAGATION FOR NUMERAL RECOGNITION

Two types of neural networks learning algorithms were created, trained, tested, and evaluated in an effort to find the appropriate neural network training method for use in numeral recognition problem. The purpose of this study was to compare the training speeds of two neural networks Backpropagation learning algorithms (Adaptive learning rate and Resilient) when exposed to ten number recognition data sets. Each algorithm was trained using ten data sets as a basic set (Boolean value), and a complex (noisy) set. The trials conducted indicated a significant difference between the two algorithms in the basic data set, with the Resilient training algorithm the neural network trained faster.The creation, training, and testing of each neural network was done using the MathWorks software package MATLAB which contains a “Neural Network Toolbox” that facilitates rapid creation, training, and testing of neural networks.MATLAB was chosen to use for learning algorithm development because this toolbox would save an enormous amount programming effort.