EvaluationΒΆ
ReX has a few basic analyses built in. These are
area
KL divergence (from a uniform distribution)
robustness to noise
insertion curve
deletion curve
total time (including analysis time)
If you call ReX with the --analyse flag, you might get something like this:
INFO:ReX:path tests/test_data/peacock.jpg, classification 84, area 0.0167, KL divergence 4.055,
robustness 0.0, insertion curve 1.06958, deletion curve 0.682799, time 8.3669
area is the size of the sufficient explanation as a ratio of the entire image.
KL divergence is the
Kullback-Liebler divergence of the
responsibility map against a uniform distribution: it measures how far the responsibility map is from flatness in bits.
Flatness usually indicates that the explanation is not local (it is distributed evenly over the image) or that
something has gone very wrong!
robustness takes the sufficient explanation and superimposes it over random backgrounds. It measures how
often the sufficient explanation manages to maintain its classification in the face of noise. When an explanation is small, robuestness is usually low.
In this case, the explanation is just over 1% of the image, and
robustness in the face of noise is 0.0. You might try increasing the --confidence threshold, or looking for a contrastive explanation
if you want something more robust.
insertion and deletion are the AUC for insertion/deletion curves (see the RISE paper).
By default, these are normalised by the original confidence of the model.
time indictates the total time spent, including analysis time.
