Avoiding the Oops with MLOps
summary
Industry stats show 87% of ML projects never leave experiments, so this talk focuses on actionable MLOps steps: use version control plus CI/CD to automate delivery of training pipelines, wire automated training into environments, and monitor for model drift (data drift vs concept drift) with statistical checks. Example: a spam filter shifts from text to attachment-heavy emails. When drift appears, trigger retraining or run scheduled retrains; tradeoffs include delayed fixes versus unnecessary compute.
AI-written summary · about
chapters (12)
- 0:00Intro
- 0:23Whats the Oops
- 1:12Development and Deployment
- 1:47Levels of Maturity
- 3:32Continuous Integration
- 4:32Model Drift
- 4:58Data Drift
- 5:55Concept Drift
- 6:47Longevity Maintenance
- 8:31Retraining
- 9:17Conclusion
- 9:45Outro
related talks
43mNavigating the MLOps Journey: Key Considerations for Successful Implementations
16mMLOps — Understanding and Fighting Model Degradation in Production
13mMLOps: From AI Research to Production-Ready Systems
20mBringing CI/CD Practices to Machine Learning with MLOps
25mBuilding Scalable MLOps Pipelines That Actually Work
13mMLOps at Scale
47mMachine Learning in Production: An Intro to MLOps.
01mMLOps: Applying DevOps to ML