Efficient Self-supervised Vision Transformers for Representation Learning