Disentangling Uncertainties by Learning Compressed Data Representation