Toward Robust Real-World Inference: A New Perspective on Explanation-Based Learning