Jumpy Recurrent Neural Networks

Recurrent neural networks (RNNs) can learn complex, long-range structure in time series data simply by predicting one point at a time. Because of this ability, they have enjoyed widespread adoption in commercial and academic contexts. Yet RNNs have a fundamental limitation: they represent time as a series of discrete, uniform time steps. As a result, they force a tradeoff between temporal resolution and the computational expense of predicting far into the future. To resolve this tension, we propose a Jumpy RNN model which does not predict state transitions over uniform intervals of time. Instead, it predicts a sequence of linear dynamics functions in latent space and intervals of time over which their predictions can be expected to be accurate. This structure enables our model to jump over long time intervals while retaining the ability to produce fine-grained or continuous-time predictions when necessary. In simple physics simulations, our model can skip over long spans of predictable motion and focus on key events such as collisions between two balls. On a set of physics tasks including coordinate and pixel observations of a small-scale billiards environment, our model matches the performance of a baseline RNN while using a fifth of the compute. When used for model-based planning, our method matches a baseline RNN while using half the compute.