frozen-feature-extractors

This paper reports that deep reinforcement learning agents using frozen, randomly initialized CNN feature extractors spontaneously develop extremely sparse fully-connected representations, compressing task-relevant information through very few neurons without any sparsity-inducing objective.