A Neural Network Model for Learning Data Stream with Multiple Class Labels
T. Takata, Seiichi Ozawa · 2011
In this paper, we extend the sequential multitask learning model called Resource Allocating Network for Multi-Task Pattern Recognition (RAN-MTPR) proposed by Nishikawa et al. such that it can learn a training sample with multiple class labels which are originated from different lassification tasks. Here, we assume that no task information is given for training samples. Therefore, the extended RAN-MTPR has to allocate multiple class labels to appropriate tasks under unsupervised settings. This is carried out based on the prediction errors in the output sections, and the most probable task is selected from the output section with a minimum error. Through the computer simulations using the ORL face dataset, we show that the extended RAN-MTPR works well as a multitask learning model.