More Design Patterns For Machine Learning Systems
Eugene Yan
The article presents design patterns for machine learning systems including processing raw data once, human-in-the-loop annotation, data augmentation, hard negative mining, and problem reframing, with examples from companies like Meta, DoorDash, and Uber. A specific finding shows that LLMs like GPT-3.5-turbo outperform crowdsourced workers on annotation tasks at 5% of the cost, and optimal training often uses a blend of hard and easy examples at ratios like 100:1. These patterns help reduce redundancy, improve model generalization, and increase training efficiency across machine learning systems.
Why it matters
9 patterns including HITL, hard mining, reframing, cascade, data flywheel, business rules layer, and more.
