We propose an adaptive subset selection framework for supervised learning, where submodular functions quantify sample difficulty to guide curriculum learning.