Back to the Future: Cycle Encoding Prediction for Self-supervised Video Representation Learning.