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