16. Transfer Learning

Why does Transfer Learning Work?

In the example above, we are leveraging the weights of an existing classification network (one that classifies cats, dogs, rabbits, and horses); we're keeping the first few layers of this trained network then adding and tre-training a final classification layer to create a new classification network that identifies humans and birds.

But why are these first few layers useful?

Well, in the case of image classification, the first layers in a network are responsible for identifying general patterns in shape and color, such as simple lines and textures, that distinguish different classes of images; later layers identify more complex features such as eyes, ears, and feet. The earlier layers, because they identify simple patterns, will be relevant in any classification case (there will always be simple lines and colors that we can use to identofy different objects)! And that is why you can transfer learnings from earlier layers of one network to another.