Pattern Recognition of Arabic/Persian Handwritten Digits using Adaptive Boosting, Neural Networks and Deep Boltzmann Machines

Abstract

Classification of handwritten numerals has captured the attention of the statistical and machine learning community. A common statistical approach is ensemble learning, which combines basic classifiers to produce one powerful predictor. Another popular method is neural networks, which is a non-linear two stage statistical model inspired by the biological neural networks. In this thesis, neural networks (nnet), adaptive boosting (AdaBoost), and deep Boltzmann machine (DBM) algorithms are tested for Arabic/Persian handwritten digits, and their recognition performance is then compared.