Testing Noise Assumptions of Learning Algorithms