Attentive Cross-Modal Connections for Learning Multimodal Representations from Wearable Signals for Affect Recognition