by Michael Munn
First published 2020
Machine learning engineers face the same problems repeatedly. Training models fails in predictable ways. Data representation choices break systems. Deployment strategies collapse under real conditions. Valliappa Lakshmanan, Sara Robinson, and Michael Munn have seen these failures hundreds of times at Google Cloud. They collected thirty proven solutions that work across different ML contexts. Each pattern tackles a specific recurring challenge. The authors explain the underlying problem, present multiple solution approaches, and recommend which remedy fits different situations. The patterns cover data representation techniques like embeddings and feature crosses. They address model selection for particular problem types. They show how to build training loops with checkpoints and hyperparameter tuning. They detail deployment strategies for scalable systems that handle retraining and updates. They provide methods for interpreting predictions and ensuring fair treatment of users. The book transforms expert experience into actionable guidance for common ML engineering obstacles.
Genres: non-fiction, technology, artificial-intelligence, computer-science, engineering, algorithms
Vibes: thought-provoking
Tropes: expert-advice
400 pages · Paperback · O'Reilly Media