Constellation: Learning relational abstractions over objects for compositional imagination

Density-based bonuses on learned representations for reward-free exploration in deep reinforcement learning

Improved predictives for neural networks through linearization