Super Data Science: ML & AI Podcast with Jon Krohn 661: Designing Machine Learning Systems
Mar 14, 2023
Chip Huyen, author of 'Designing Machine Learning Systems', shares insights on real-time machine learning, training data, feature engineering, and productivity tips. She discusses aligning business intent with ML, secrets to good features, optimizing productivity, and the importance of real-time ML applications.
AI Snips
Chapters
Transcript
Episode notes
Netflix Winning Model Never Deployed
- At Netflix an experimentally winning solution was never deployed because it was too complex to operationalize.
- Chip used that experience to learn the gap between research prototypes and deployable systems.
Tune Model Update Frequency To Your Domain
- Measure how delay in model updates affects performance and choose update cadence per application.
- Some systems (e.g., recommender click predictions) require hourly or sub-hourly updates to avoid significant metric decay.
Why ML Appears Magical
- ML models feel like magic because we reverse the usual software design: we give input and output and let the system discover the function.
- Large parameter counts make understanding internals infeasible, creating the perception of magic.