An Occam's Razor View on Learning Audiovisual Emotion Recognition with Small Training Sets