ADAPTIVE MIXTURE LEARNING IN DYNAMIC SYSTEM

摘要

PPROBLEM TO BE SOLVED: To provide an improved method and a filter with respect to conventional data modeling technique for modeling real-time data such as an audio stream and a video. PSOLUTION: An adaptive learning rate schedule is used for an on-line Gaussian mixture learning model for dynamic data, in order to achieve rapid convergence while keeping adaptive possibility in the model after convergence. An experimental result shows an unexpected improvement with respect to modeling precision by using the adaptive learning model. PCOPYRIGHT: (C)2004,JPO