Text this: Reinforcement learning on slow features of high-dimensional input streams.