Towards Solving Inefficiency of Self-supervised Representation Learning