Challenges in incorporating ML models
summary
A machine learning integration project hits production roadblocks: a model that needs under 50 ms latency and trains for two weeks on four GPU machines, achieves near 98% coverage on first pages yet stalls when product metrics, error-handling and front-end integration aren’t solved, and the project is ultimately scrapped in favor of an external vendor. Recommended tactics are pragmatic: ship business KPIs (time saved, throughput, cost), use TensorFlow 1.x for stability, narrow the distribution or build modular pipelines, and add robust backward-compatible rollout and error workflows.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:40Project from hell
- 3:40Learning the code
- 7:20What is ML
- 8:55Training
- 9:55Errors
- 13:30Getting through the door
- 16:29Trust and love
- 19:55Generalization vs overfitting
- 27:21Pipeline
- 33:53Conclusions
related talks
30mChallenges of machine learning: The new ML lifecycle
31mBuilding the Business case for ML Platforms
21mWhy is Production ML so Hard?
27mKeys to Building Machine Learning Systems
29mDemystifying Machine Learning in Production: Reasoning about a Large-Scale ML Platform
22mAccelerating ML impact
01h 05mMachine Learning Models and Productionization
41mInfrastructure for ML Applications