NeurIPS 2021: Learning State Representations from Random Deep Action-conditional Predictions - View it on GitHub