An auxiliary classifier is an extra prediction head attached to an intermediate layer of a deep network. In GoogLeNet, softmax predictions from two intermediate Inception modules receive losses against the ground truth, helping prevent the middle layers from dying under vanishing gradients. section 11.2.2
GoogLeNet loss
GoogLeNet combines the main classification loss with the two auxiliary losses as a weighted sum: total loss = real loss + weighted auxiliary losses. The classifiers are attached within a 22-layer network, counted as 27 layers when pooling layers are included. section 11.2.2
Figure 11.8 Inception_v1 architecture
Use the Inception_v1 architecture diagram to locate the intermediate modules that feed auxiliary classifiers and the final classifier. section 11.2.2
