Provably Efficient Exploration for Reinforcement Learning Using Unsupervised Learning