Reviews: Visual Interaction Networks: Learning a Physics Simulator from Video

This paper presents a general-purpose model based on convolutional networks and recurrent neural networks with an interaction network to predict future states from raw visual observations. The use case is compelling and the models demonstrate excellent performance in predicting future states and also demonstrating good performance with noisy backgrounds and springs with invisibility and springs and billiards with variable mass. MAJOR - in the section training procedure 3.3, please clarify the sequence of 8 unseen future states. I read from 3.1 that the training sequence is 14 frames. Are these next 8 unseen future state frames the next 8 frames after the training sequence of 14 frames or are these randomly selected frames from future time? - apologies but the actual setup of the tests is unclear to me from the descriptions.