Python Deep Learning - Second Edition

Inception networks

Inception networks (https://www.cs.unc.edu/~wliu/papers/GoogLeNet.pdf by Szegedy et al.) were introduced in 2014, when they won the ImageNet challenge of that year (there seems to be a pattern here). Since then, the authors have released multiple improvements (versions) of the architecture.

Fun fact: the name inception comes in part from the “We need to go deeper” internet meme, related to the movie Inception.

The idea behind inception networks starts from the basic premise that the objects in an image have different scales. A distant object might take up a small region of the image, but the same object, once nearer, might take up the majority of the image. This presents a difficulty for standard CNNs, where the neurons ...