Study On Hybrid Incremental Learning Based On Intelligent Human-computer Interaction Systems

For human beings, it is an important process to learn new knowledge incrementally because of environment changes in their lives.Similarly, the environment changes also requires the ability of self-learning for intelligent interaction systems. In real-world scenarios,the environment are open-ended and dynamic: the object categories and object instances are constantly updated, and even the meaning of the concept is evolving. Therefore, learning new knowledge incrementally in a hybrid way is very important for intelligent Human-Computer Interaction (HCI) system. On the another side, the configuration of smaller units of information into large coordinated ones might be important in many processes of human perception, learning and cognition. Therefore, HCI systems require a mechanism to automatically learn new information incrementally by visual recognition and make associations between concepts. As a very special yet important case of object recognition, hand-held object recognition plays an important role in HCI system. And our study is based on hand-held object recognition system.In this paper, we present a Hybrid Incremental Learning (HIL)method based on Support Vector Machine (SVM), which can learn new class by adding new classification-plane and learn new samples by modifying learned classification-planes. In this paper, we also combine a knowledge graph (KG) and incremental learning to learn new information and organize learned information systematically and efficiently. The hybrid incremental learning can automatically learn new objects by adding new nodes to the KG and can improve the recognition capability of classifiers related to existing nodes. As a result, the hybrid incremental learning method can propagate the new information through the KG, as a knowledge ripple effect which updates more than one classifiers related with different nodes.This paper validates the effectiveness of the algorithm and the framework on two datasets. The experimental results proves and the algorithm is able to learn new information based on the existing model,and the framework make the learned information related to each other,and make the learning process more effective.