A Deep One-Pass Learning based on Pre-Training Weights for Smartphone-Based Recognition of Human Activities and Postural Transitions

Abstract

We describe a new deep learning model – Deep One-Pass Learning (DOPL) for Smartphone-Based Recognition of Human Activities and Postural Transitions based on the Pre-Trained Weights, DOPL consists of several stacked convolutional layers to learn the features of the input and is able to learn features automatically, followed by the Extreme gradient boosting (XGBoost) as the last layer for predicting the class labels. DOPL is much faster in the training phase, because the input weights are optimal weights from the Pre-Trained weights module and it does not have to re-adjust weights repeatedly. Further, we replaced the final fully connected layer with XGBoost to increase predictive efficiency. In the worst case, our model with demonstrated an accuracy of 99.2% for the smartphone sensors database data, which was significantly better than CNN or XGBoost alone as well as several other models assessed.