Now, you might wonder why we want to extract the features from an intermediate layer in a CNN. The key intuition is that: as the network learns to classify images into categories, each layer learns to identify the features that are necessary to do the final classification.

Lower layers identify lower order features such as colors and edges, and higher layers compose these lower order features into higher order features such as shapes or objects. Hence the intermediate layer has the capability to extract important features from an image, and these features are more likely to help with different kinds of classification.

This has multiple advantages. First, we can rely on publicly available large-scale training and transfer this learning to novel domains. Second, we can save time for expensive large training. Third, we can provide reasonable solutions even when we don't have a large number of training examples for our domain. We also get a good starting network shape for the task at hand, instead of guessing it.