Conjugate prior penalized learning of Gaussian mixture models for EMG pattern recognition

This paper proposes a new learning method for a Gaussian mixture model (GMM). First, a traditional maximum a posterior (MAP) parameter estimate is used to achieve regularization based on conjugate priors. Next, a model order selection criterion is derived from Bayesian-Laplace approaches such that the conjugate prior distribution can be used to measure the uncertainty in the estimated parameters. As a result, the proposed learning method avoids the possibility of convergence toward local minima in the parameter space, and is also capable of selecting the optimal order for a GMM using an additional complexity penalty for the prior distribution. The proposed method is applied to electromyogram (EMG) pattern recognition for controlling a multifunction myoelectric hand, and experiments conducted to recognize nine kinds of hand motion from EMG signals for ten subjects. In conclusion, the proposed learning method effectively estimated the change of feature vectors according to the subject and the GMM classifier demonstrated a high recognition accuracy.