Figure 3.

Vectorized computation. The three neural networks in this diagram are used sequentially, ordered by their associated step sizes from large to small. For each network, we stack all currently existing states and step forward (in the beginning, we only use the initial state), resulting in vectorized computations. These newly generated states are further fed to the next neural network in queue. Once we finish using all networks, the states will be rearranged in terms of chronological order and intermediate time steps will be obtained via interpolation. (Online version in colour.)