Machine Learning Design Patterns

Chapter 8. Connected Patterns

We set out to create a catalog of machine learning design patterns, solutions to recurring problems when designing, training, and deploying machine learning models and pipelines. In this chapter, we provide a quick reference to this inventory of patterns.

We organized the patterns in the book in terms of where they would be used in a typical ML workflow. Thus, we had a chapter on input representation and another on model selection. We then discussed patterns that modify the typical training loop and make inference more resilient. We ended with patterns that promote a responsible use of ML systems. This is akin to organizing a recipe book with separate sections on appetizers, soups, entrees, and desserts. Such an organization, however, can make it hard to determine when to choose which soup and what desserts go well with some entree. Therefore, in this chapter, we also draw out how the patterns are related to one another. Finally, we also put together “meal plans” by discussing how the patterns interact for common categories of ML tasks.

Patterns Reference

We’ve discussed a lot of different design patterns and how they can be used to address common challenges that arise in machine learning. Here is a summary.

Chapter	Design pattern	Problem solved	Solution
Data Representation	Hashed Feature	Problems associated with categorical features such as incomplete vocabulary, model size due to cardinality, and cold start.	Bucket a deterministic and portable ...