Commodity Machine Learning
summary
Argues machine learning is becoming a commodity by showing concrete wins: psyched learn plus pandas and IPython Notebook let common supervised tasks be solved in a few lines, with ~500–600 papers citing the library. Recommends single big EC2 box (32 cores, 256 GB) for most workloads and notes Kindle language models as a rare cluster case. Highlights Spark MLlib and Blaze for scale, and demonstrates Bayesian hyperparameter search / AutoML prototypes that actually try many models and return dashboards.
AI-written summary · about
chapters (15)
- 0:00Commodity ML introduction
- 3:46Current ML tools
- 9:40Python ML limits
- 16:49Single-machine sufficiency
- 21:10Need for AutoML
- 27:41When distributed ML
- 29:51Automation over scalability
- 30:17Q&A: Data cleaning
- 32:07Q&A: Feature engineering
- 33:34Q&A: ML commodity?
- 34:38Q&A: ML methods
- 36:44Q&A: Hyperparameters & overfitting
- 38:31Q&A: Model interpretability
- 40:38Q&A: AutoML risks
- 41:55Closing remarks
related talks
22mReal-world applications of machine learning: Beyond the hype
35mIntelligent Ruby: Getting Started with Machine Learning
40mCloud-based Machine Learning Offerings - Beginner's Perspective
04mThe machine-learning renaissance
27mA Look Inside Applied Machine Learning
44mPragmatic Machine Learning in the Cloud
26mWhy Machine Learning Algorithms Fall Short...
17mMachine Learning at the Speed of Light with Spark