In the previous chapter, we learned that a CNN consists of several layers. We also studied different CNN architectures, tuned different hyperparameters, and identified values for stride, window size, and padding. Then we chose a correct loss function and optimized it. We trained this architecture with a large volume of images. So, the question here is, how do we make use of this knowledge with a different dataset? Instead of building a CNN architecture and training it from scratch, it is possible to take an existing pre-trained network and adapt it to a new and different dataset through a technique called transfer learning. We can do so through feature extraction and fine tuning.
Transfer learning is the process of copying knowledge from an already trained network to a new network to solve similar problems.
In this chapter, we will cover the following...