Learning Invarting Representations for Reinforcement Learning without Reconstruction