August 2018
Intermediate to advanced
522 pages
12h 45m
English
If we have a target probability distribution p(x), which is approximated by another distribution q(x), a useful measure is cross-entropy between p and q (we are using the discrete definition as our problems must be solved by using numerical computations):

If the logarithm base is 2, it measures the number of bits requested to decode an event drawn from P when using a code optimized for Q. In many machine learning problems, we have a source distribution and we need to train an estimator to be able to correctly identify the class of a sample. If the error is null, P = Q and the cross-entropy is minimum (corresponding ...
Read now
Unlock full access