GoogLeNet was a breakthrough convolutional neural network architecture, applied for image classification.
Architecture
Architecture’s primary motivation: going deeper. The original GoogLeNet (from 2014) had 22 convolutional layers.
Intermediate classifiers
One of the key problems with deeper models is the vanishing gradient problem, because of successive multiplications through layers. The idea is simple: classifiers are used in intermediate steps of the model (purple to red branches), and the losses in the classifier mean that all parts of the network are properly covered.
The total loss is a combination of the intermediate and final losses.