Training of pre-trained models is done on many comprehensive image classification problems. The convolutional layers act as a feature extractor, and the fully connected (FC) layers act as classifiers, as shown in the following diagram, in the context of cat vs. dog image classification with a conv net:
Since the standard models, such as VGG-16/19, are quite large and are trained on many images, they are capable of learning many different features for different classes. We can simply reuse the convolutional layers as a feature extractor that learns low and high level image features, and train only the FC layer weights ...
