Hierarchical learning of primitives using neurodynamic model

Abstract

Primitives are essential features in human’s recognition capability. In this paper, we propose a method to form representations of motion primitives within a neural network model to learn sequences that consist of a combination of primitives. The training model is based on human development phenomena, namely motionese, where human infants learn simple motions first, and tend to learn complex motions as their skills develop. A neurodynamical model, Multiple Timescale Recurrent Neural Network (MTRNN), is used for the dynamics learning model. Neurons in MTRNN are composed hierarchically to learn different levels of information. The proposed method first trains the model using basic simple motions with neurons learning low level data. After training converges, neurons learning high level data are attached to the model, and complex motions are used to train the model. Experiments and analysis with drawing data show the models capability to form representations of primitives within the model.