Analysis of Supervised Learning Methods on Artificial Neural Networks

Abstract

Artificial intelligence applications are now highly developed, where hardware is given the ability to think and act like humans. One method for building artificial intelligence applications is artificial neural networks (ANN). ANN is currently highly developed in areas of application that start as diverse as classification, prediction, games, robots, IoT, object/sound recognition, and so on. Unlike the human brain, which is designed to solve many problems, ANN can only solve a specific problem. To make ANN intelligent, learning methods are needed to train ANN to be able to solve certain problems with a good degree of accuracy. There are 3 learning methods commonly used to make ANN intelligent, are perceptron, backpropagation, and extreme learning method. Perceptron is the simplest learning method for the simplest ANN structure. Backpropagation was found to answer the weaknesses of perceptron which cannot solve nonlinear problems. And finally ELM exists to overcome the problem of long computational time in perceptron and backpropagation learning. This research will conduct an analysis of the three ANN learning methods using several problems. Problems used include AND logic, OR logic, XOR logic, iris datasets and MNIST digit handwritten digits. Because the ANN learning method is based on random parameters, each case will be trained 10 times for each method. After testing it is concluded that ELM has a very good speed with a pretty good degree of accuracy.