Self-reflective learning of invariants in human-artifact interactions using recurrent neural network

Human bodily motions are effectively used to communicate, and the ability to read the intentions behind those is essentially important for the machine system to collaborate with the human. If we use behaviors as medium of communication with the machine, the machine system should be able to construct meanings from them. In this paper, semiosis and related topics of symbol grounding are reviewed, and motion understanding is discussed in terms of that framework. Then, a method for extracting meanings from a series of human bodily motion is presented using a recurrent neural network (RNN). Finally we discuss about what the RNN learning implies with respect to semiosis