How Can We Trust Machine Learning? Exploration, Evaluation...
summary
Companies use GraphLab Create to train a recommender in minutes and ship it to AWS as a real‑time microservice via Predictive Services (demo uses a 3‑node cluster pulling models from S3 and serving <7 ms responses). University of Washington research builds explanations by local perturbation and fitting simple models to highlight important features; on 20 Newsgroups this reveals spurious signals (email addresses) and removing them drops accuracy to ~57%. Mechanical Turk users improve models using those explanations.
AI-written summary · about
chapters (19)
- 0:00Intro
- 0:40Every innovative successful application will use machine learning
- 0:59Intelligent applications can provide tremendous value
- 1:30Three core blockers
- 3:04Data
- 4:16Our Goal
- 4:43Agile Production Path
- 6:04Demo
- 11:56Trusting ML Models
- 12:51Live User Study
- 14:27Predicted Wolf
- 16:25Understanding Why
- 17:06Improving Quality
- 17:43Examples
- 19:14Building Trust
- 19:57Churn Prediction
- 20:48Online Courses
- 21:21Machine Learning Applications
- 22:12Summary
related talks
46mKeynote: How Can We Trust Machine Learning?
30mOperationalizing Machine Learning
30mPython based predictive analytics with GraphLab Create
44mPragmatic Machine Learning in the Cloud
01h 30mBecome a data-driven Organization with Machine Learning
29mMachine Learning in Production: What to do after you’ve trained the models?
22mReal-world applications of machine learning: Beyond the hype
41mRecommendations for Building Machine Learning Software