Changing Realities - The Final Frontier for Machine Learning?
summary
Claim: real-world concept drift breaks static models and demands either continuous online learning or engineered features. Built an occupancy predictor for buses, tries online learning with Apache MOA and trained 20 algorithms per line, sees day‑three adjustments but fails in production because only 10–20% of occupancy data exists, data holes and shifting feature importances break automatic adaptation. Alternative succeeds: heavy engineering of adaptive calibration features (relative week‑over‑week normals, percentage-of-total, normalization), but it requires significant human effort.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 1:17Adaptive AI overview
- 3:07Public transport occupancy
- 8:04Traditional ML limits
- 12:45Changing realities impact
- 13:47Concept drift effects
- 17:19Handling concept drift
- 18:47Online learning training
- 23:20Adaptive calibration strategy
- 28:36Key takeaways
- 31:05Q&A automation
- 35:44Q&A non-time-series
- 38:55Closing remarks
related talks
41mMachine Learning Models in a Dynamic Environment
30mReconceptualizing Machine Learning for the Real-time World
18mReliable Maintenance of Machine Learning Models
24mChallenges in building Machine Learning models in production
48mPrinciples and Practices of Building Machine Learning Systems
51mKeynote: Fault Tolerant Machine Learning Operations
39mMachine Learning Engineering
32mReliable Machine Learning