October 2017
Beginner to intermediate
270 pages
7h
English
In December 2015, with the paper Rethinking the Inception Architecture for Computer Vision, Google Research released a new iteration of the Inception architecture.
The internal covariance shift problem
One of the main problems of the original GoogLenet was training instability. As we saw earlier, input normalization consisted basically of centering all the input values on zero and dividing its value by the standard deviation in order to get a good baseline for the gradients of the backpropagations.
What occurs during the training of really large datasets is that after a number of training examples, the different value osculations begin to amplify the mean parameter value, like in a resonance phenomenon. ...
Read now
Unlock full access