Learning Affective Features With a Hybrid Deep Model for Audio–Visual Emotion Recognition