Neural networks that perform classification(predict the class of an input) produce a vector of C
raw numbers for every input,
where C
is the total number of classes, for example 10 for MNIST. This vector is called a logit.
If we feed an MNIST image for digit 3 to a classification model, it may produce the a logit vector like this (top row added to show the class ID's:
0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 |
---|---|---|---|---|---|---|---|---|---|
-121.2 | -212.3 | 81.1 | 171.1 | -55.0 | 132.5 | -13.2 | 63.5 | 99.2 | -10.9 |