Acquiring New Categories by Self Data Gathering with Bayesian Attractor Model

Abstract

In the IoT, AI should automatically learn the new categories when the environment changes caused by moving, adding, or removing the devices. In literature, several methods are proposed for learning in a new environment such as selftraining and transfer learning. However, there are yet some issues to learn the new category related to small data for the new category and noise existence. On the other hand, noise-tolerant models have been developed such as the Bayesian Attractor Model (BAM) which is a cognitive model for decision making under uncertainty of information. In this paper, we develop a system to learn the new category automatically by extending the BAM. Using the BAM as a classifier, the system makes the temporal classifier with new observation and collects the training data for the new category. Through the evaluation, we show that our method is superior to the neural network approach in the scenario of learning new categories.