Existing works on action recognition rely on two separate stages: (1) designing hand-designed features or learning features from video data; (2) classifying features using a classifier such as SVM or AdaBoost. Motivated by two observations: (1) independent component analysis (ICA) is capable of enco