Video Prediction Efficient and Information-Preserving Future Frame Prediction and Beyond [ICLR 2020]. Wei Yu, Yichao Lu, Steve Easterbrook, Sanja Fidler. reduce origin pixel-wise prediction computational take advantage of 3D CNN info introduce a two-way autoencoder which can reduce memory demand and computational cost, but also no information loss in theoretically guaranteed. propose reversible predictive modules(RPM).