A Dropconnect Deep Computation Model for Highly Heterogeneous Data Feature Learning in Mobile Sensing Networks

Deep computation model, as a tensor deep learning model, outperforms multi-modal deep learning models for feature learning on heterogenous data. However, deep computation model is limited in generalization to small heterogeneous data sets since it typically requires many training objects to learn the parameters. In this article, we propose a dropconnect deep computation model (DDCM) for highly heterogeneous data feature learning in mobile sensing networks. Specifically, the dropconnect technique is used to generalize the large fully-connected layers in the deep computation model for small heterogeneous data sets. Furthermore, the rectified linear units (ReLU) are used as the activation function to reduce computation and prevent overfitting. Finally, we compare the classification accuracy and execution time for learning the parameters between our model and the traditional deep computation model on two highly heterogeneous data sets. Results illustrate that our model achieves 2 percent higher classification accuracy and performs more efficiently than the deep computation model, proving the potential of our proposed model for highly heterogeneous data learning in mobile sensing networks.