Extracting Explanations, Justification, and Uncertainty from Black-Box Deep Neural Networks