Suppressing Static Visual Cues via Normalizing Flows for Self-Supervised Video Representation Learning.