Incremental Learning, Recognition, and Generation of Time-Series Patterns Based on Self-Organizing Segmentation

Abstract

We segments and symbolizes image information on a series of human behavior as an aggregate unit of motions in a self-organizing manner and proposes a system that recognizes the entire behavior as a symbol string. This system symbolizes the motion unit incrementally and also generates motion from a symbol. To implement the system, we used a mixture of experts with a non-monotonous recurrent neural network used as the expert and our own DP matching method. In addition, our proposal makes not only teacher-labeled patterns, but also teacher-unlabeled patterns available for learning. By using this function, we proposed semi-supervised learning using our proposal in this paper. We verified the evaluation of the effectiveness of our proposal and semi-supervised learning function by two experiments using moving images including seven gestures.