There is a minor error in the Proceedings version of the paper, which has been corrected in the web version. In Section 3, just below Figure 6, we wrote:
INCORRECT | For a globally stable metric, we require the average change and the standard deviation of change to be within the above threshold for all inputs in the training set. |
This is wrong. It must instead read:
CORRECT | For a globally stable metric, we require the average change and the standard deviation of change to be within the above threshold for at least 40% of the inputs in the training set. |