Weight Initialization-based Partial Training Algorithm

The deep learning or deep neural network (DNN) currently provide the best solution to many problems in classification. However, DNN requires much more time and many calculations than existing other classification algorithms. In addition, if new features and classes are added into existing model, it should re-learn all of data set in order to apply the new features and classes into learned model. And it consumes a lot of computing powers and time due to the characteristics of DNN. To resolve this problem, in this paper, we proposed partial-learning algorithm. The proposed algorithm allows to learn new features and classes without whole re-learning of dataset. For this, the proposed algorithm creates latent matrices for new weight parameters through extension of weight matrices in the existing model. Finally, we shown possibility of the proposed algorithm via experiment.