Summary

Machine learning and traditional software are often perceived as two alternatives. However, they are more like siblings – one cannot function without the other. Machine learning models are very good at solving constrained problems, but they require traditional software for data collection, preparation, and presentation.

The probabilistic nature of machine learning models requires additional elements to make them useful in the context of complete software products. Therefore, we need to embrace this nature and use it to our advantage. Even for safety-critical systems, we could (and should) use machine learning when we know how to design safety mechanisms to prevent hazardous consequences.

In this chapter, we explored the differences between machine learning software and traditional software while focusing on how to design software that can contain both parts. We also showed that there is much more to machine learning software than just training, testing, and evaluating the model – we showed that rigorous testing makes sense and is necessary for deploying reliable software.

Now, it is time to move on to the next chapter, where we’ll open up the black box of machine learning software and explore what we need to develop a complete machine learning software product – starting from data acquisition and ending with user interaction.