Running Machine Learning in Production - a Journey to Success
summary
Shows concrete wins and failures: trains a 242k-parameter CNN on the German traffic-sign set in Google Colab, raising test accuracy from ~93% to ~97% and AUC from ~0.99 to ~0.999 by increasing training samples; labels matter — a packaged-goods project uses 100–120k labeled images with human relabeling agreement 97–99% (inter-annotator 95–98%); prototype hockey detection uses YOLOv4, hitting 2.95 fps initially, later 24 fps (vs. claimed 65 fps); emphasizes runtime, latency, power and choosing simpler models when data or explainability matters.
AI-written summary · about
chapters (9)
- 0:00Introduction and journey
- 2:13Data is king in production
- 8:29Labeling challenges and data quality
- 12:22Demo: More data improves accuracy
- 19:08Model selection beyond accuracy
- 23:42Demo: Decision trees for explainability
- 27:21Know your runtime platform
- 30:31Key lessons for ML in production
- 32:01Q&A on missing data
related talks
41mBeyond the Algorithm: What Makes Machine Learning Work?
40mHow to Implement a Production Ready Deep Learning Model in 12 Weeks
33mOvercoming the Barriers to Deep Learning in Production
30mMachine Learning: Beyond the Tutorial
36mApplied Machine Learning Playbook: From Idea to Production
24mChallenges in building Machine Learning models in production
25mContinuous ML Applications in Production
40mMachines are Learning