Keynote: How Can We Trust Machine Learning?
summary
Argues a concrete path to trust: uses GraphLab Create and deployable microservices to take models from laptop to real-time AWS services, shows a deep network that finds visually similar dresses and a text-analysis tool that surfaces aspects of baby‑monitor reviews, reports a surprising drop from ~95% to 57% accuracy on 20 Newsgroups after removing leaked email features and then mechanical turkers improve the model beyond the cleaned gold standard in three rounds, and claims sub‑5 ms prediction latency and SFrame out‑of‑core scaling to huge graphs.
AI-written summary · about
chapters (10)
- 0:00Introduction and journey
- 3:02Building intelligent applications
- 10:49Real-world applications and demos
- 14:01Demo of recommender system
- 21:44The trust challenge in ML
- 28:13Explaining model predictions
- 32:08Tools for scalable ML
- 35:47Online education in ML
- 39:41Summary and mission
- 40:02Q&A session
related talks
23mHow Can We Trust Machine Learning? Exploration, Evaluation...
54mKeynote: Can you trust your (large language) model?
39mKeynote 3: How did you know? Explaining Black Box Model Predictions in Python
51mKeynote: Fault Tolerant Machine Learning Operations
22mKeynote talk - The future of machine learning & JS
49mKeynote 4: Here Come The Robots - Python and Machine Learning
30mMachine Learning: Beyond the Tutorial
37mKeynote: Ethical Machine Learning: Creating Fair Models in an Unjust World