Looks interesting so I jumped to the experiments. The bouncing balls example as in Fig 3 looks rather unconvincing. When collision occurred in +12, LSTM has no

Looks interesting so I jumped to the experiments. The bouncing balls example as in Fig 3 looks rather unconvincing. When collision occurred in +12, LSTM has no problem learning the bounce-off, its trajectory is a little off compared with RIM. This reminds me of anther paper[1] which addresses a very similar problem. [1] Srivastava, N., Mansimov, E., & Salakhudinov, R. (2015, June). Unsupervised learning of video representations using lstms. In International conference on machine learning (pp. 843-852). EDIT : They use exactly the same bouncing MNIST dataset as in [1]. I'm surprised they didn't cite this paper.