The first row shows the testing confusion pair, the second row shows the top 10 harmful training identified by IF, and the third row is the harmful training identified by EIF.

Visually, given a pair of confusing testing samples, EIF identifies top harmful training samples (1) more visually similar to the testing pair (compared to those from IF) and (2) distributed in concentrated classes (In the first image, they concentrate on class 42. In the second image, they concentrate on class 98.)

Visually, given a pair of confusing testing samples, EIF identifies top harmful training samples (1) more visually similar to the testing pair (compared to those from IF) and (2) distributed in concentrated classes (In the first image, they concentrate on class 83. In the second image, they concentrate on class 42&45. In the third image, they concentrate on class 42.)

Visually, given a pair of confusing testing samples, EIF identifies top harmful training samples (1) more visually similar to the testing pair (compared to those from IF). One interesting observation is that: The DML model confuses two images when they have similar backgrounds and use the same clothing model. The top harmful samples given by EIF better reflect such a bias.