Towards Reliable and Efficient Representation Learning