Machine Learning, The Right Way
summary
Argues that labeled data is the bottleneck not model hype, shows simple logistic regression experiments where a single C hyperparameter shifts accuracy from ~75% to 83.7% and naive up/downsampling creates 99% training but overfits. Describes a deployed pipeline called Optimus Prime: a metadata modeling language that auto-generates feature transforms, runs on Spark/streaming/single-node, tries logistic regression and random forest per tenant, drops low-entropy features, and produces differing real-world metrics across customers.
AI-written summary · about
chapters (7)
- 0:00Introduction to AI for 99%
- 1:32The reality of data and labels
- 4:28Rethinking shallow learning
- 8:51Demonstrating model sensitivity
- 21:15Production challenges
- 24:33Core constraints in enterprise AI
- 27:45Demo of Optimus Prime architecture
related talks
42mThe Lego Model for Machine Learning
34mComplex Machine Learning Pipelines Made Easy
48mThe Missing Step in Moving from Data to Models
35mIntelligent Ruby: Getting Started with Machine Learning
38mFantastic ML apps and how to build them
04mManaging Machines
28mHindsight bias: A machine learning story
47mMachine Learning in Production: An Intro to MLOps.