Class-specific
classification rates tell a different story.
Class
|
Hits
|
Misses
|
Accuracy
|
Asian
|
0
|
26
|
0%
|
Black
|
32
|
217
|
13%
|
Hispanic
|
0
|
13
|
0%
|
White
|
1648
|
51
|
97%
|
The network's
performance on non-White faces is terribly poor.
Why?
Note that
fully 92% of the training faces
are white!
Hypothesis
The strong
imbalance in the training set means that the network is much better
conditioned for white faces than for faces from the other classes.
That
is, there simply are not enough non-White faces for the network to
form robust models of the other classes.
|