Uncertainty in Explanations

Abstract Saliency maps are popular approaches for creating post-hoc explanations of image classifier outputs. In these approaches, estimates of the relevance of each pixel to the classification output score are obtained and displayed as a saliency map that highlights important pixels. However, to be able to trust the model’s decisions and the explanation for the decision, a human analyst needs to know the confidence that model has in its predic- tions. We propose to quantify this confidence by introducing the epistemic uncertainty of the model parameters into the saliency maps using Bayesian learning and to use the uncertainty induced saliency maps to explain the classifier outputs. We investigate the utility of uncertainty in prediction scores and saliency maps to detect out-of-distribution samples and improve the trust between human analyst and the machine learning models.
  • Prudhvi Gurram (ARL)
  • Supriyo Chakraborty (IBM US)
  • Richard Tomsett (IBM UK)
  • Raghuveer Rao (ARL)
  • Lance Kaplan (ARL)
Date Sep-2020
Venue 4th Annual Fall Meeting of the DAIS ITA, 2020