Quantifying Learnability and Describability of Visual Concepts Emerging in Representation Learning