Annotation Efficiency in Multimodal Emotion Recognition with Deep Learning