In this chapter, we applied a CNN for the task of recognizing facial expressions. Using this, we could infer the emotional state of a given face. As usual, we again spent the majority of our time understanding the required input for the model and implementing the functionality to facilitate this. But, in doing so, we uncovered some important considerations when developing intelligent applications; the first is the explicit awareness of using either an end-to-end solution or a multi-step approach, with the multi-step approach being the most common one you will use.

This essentially means you, the designer and builder of intelligent applications, will be building data pipelines consisting of many models, each transforming the data in preparation for the next. This is similar to how deep networks work but provides greater flexibility. The second consideration is highlighting the availability of complementary frameworks available on iOS, in particular the Vision framework. It was used as one of the steps in our pipeline but offers a lot of convenience for common tasks, as well as a consistent workflow.

In this example, our pipeline consisted of only two steps, face detection and then emotion recognition. But we also briefly played with a feature of the Vision framework that can be used to identify facial landmarks. So, it is plausible to consider facial landmarks to train the emotional classifier rather than the raw pixels, in which case our pipeline would consist of three steps: face detection, landmark detection, and, finally, emotion recognition.

Finally, we briefly explored some use cases showing how emotion recognition could be applied; as our computers shift away from being pure tools towards being companions, being able to detect and react to the emotional state of the user will become increasingly more important. So, it's an area well worth further exploring. 

In the next chapter, we will introduce the concept of transfer learning and how we can use it to transfer styles from one image onto another.