Representations and Exploration for Deep Reinforcement Learning using Singular Value Decomposition