Chapter 10. Hyperparameter Tuning and Automated Machine Learning
In the previous chapters, we have seen how Kubeflow helps with the various phases of machine learning. But knowing what to do in each phase—whether it’s feature preparation or training or deploying models—requires some amount of expert knowledge and experimentation. According to the “no free lunch” theorem, no single model works best for every machine learning problem, therefore each model must be constructed carefully. It can be very time-consuming and expensive to fully build a highly performing model if each phase requires significant human input.
Naturally, one might wonder: is it possible to automate parts—or even the entirety—of the machine learning process? Can we reduce the amount of overhead for data scientists while still sustaining high model quality?
In machine learning, the umbrella term for solving these type of problems is automated machine learning (AutoML). It is a constantly evolving field of research, and has found its way to the industry with practical applications. AutoML seeks to simplify machine learning for experts and nonexperts alike by reducing the need for manual interaction in the more time-consuming and iterative phases of machine learning: feature engineering, model construction, and hyperparameter configuration.
In this chapter we will see how Kubeflow can be used to automate hyperparameter search and neural architecture search, two important subfields of AutoML.