The low number of parameters, as well as the network's performance, are the results of several concepts implemented by the GoogLeNet authors. We will cover the main ones in this section.
In this section, we will present only the key concepts differentiating the inception networks from the ones we introduced previously. Note that the GoogLeNet authors reapplied several other techniques that we have already covered, such as the prediction of multiple crops for each input image and the use of other image transformations during training.
