How to estimate the uncertainty of a given model is a crucial problem.
Current calibration techniques treat different classes equally and thus
implicitly assume that the distribution of training data is balanced, but
ignore the fact that real-world data often follows a long-tailed distributio