A Deep Single-Pass Learning for Recognition of Handwritten Digits

Setthanun Thongsuwan, Saichon Jaiyen

Abstract

We describe a new deep learning model - Deep Single-Pass Learning (DSPL) that can learn a data set only a single pass for recognition and prediction with high accuracy of the optical recognition of handwritten digits problem. DSPL consists of several stacked convolutional layers to learn features automatically and Extreme gradient boosting (XGBoost) is set the last layer for predicting the class labels. The learning time is O(Lc 2mnpq) + O(x(Kt + log B)), which is less than the learning time of a deep learning - Convolutional Neural Networks (CNNs). The network is no need for iteration to re-adjust the weight during the feature learning process. The results of the experiments in the test set show that our model handle the problem well and provides better accuracy than other models i.e. CNNs, XGBoost, LR, ETC, GBC, RFC, GNB, and DTC, including MLP and SVC families, with that DSPL provides 99.98% accuracy.