Handwritten Digit Recognition with Pattern Transformations and Neural Network Averaging

Abstract

Recently there has been a considerable improvement in applications related with isolated handwritten digit and letter recognition supported on the use of deep and convolutional neural networks and other combinations which make use of ensemble averaging. The proposal of the present work is based on a relatively modest sized Neural Network trained with standard Back Propagation and combined with a set of input pattern transformations. Applying ensemble averaging on the trained Neural Networks gives an encouraging error rate of 0.34% measured on the MNIST dataset.