Designing for Guidance in Machine Learning
summary
Real examples and numbers: Python code uses MNIST handwritten digits, TensorFlow yields higher raw accuracy but is a black box, while engineered features plus scikit-learn decision trees reach over 90% after rotation augmentation. Image pipeline uses skeletonize to one-pixel-wide structure, OpenCV for probabilistic Hough line transform and corner detection, then counts horizontals and quadrant corners. Guidance maps the model’s predicted target minus current input to prose/UI hints and requires logging user edits as a learning loop.
AI-written summary · about
chapters (9)
- 0:00Designing ML systems for guidance
- 3:29Types of ML systems
- 11:24Human-understandable features
- 17:49Refining guidance with behavior
- 21:48Job listing optimization example
- 24:05Q&A on guidance systems
- 31:34Explainability in ML models
- 34:26Adoption of guidance-based tools
- 37:33Scaling learning loops
related talks
45mApplied machine learning: a few lessons I learned the hard way
30mMachine Learning: Beyond the Tutorial
01h 06mDeep Learning with TensorFlow
55mTensorflow: Learning Machines
34mML from a software engineer's perspective
40mMachines are Learning
44mHow to Improve UX with Machine Learning: Insights and learnings until now
38mDesigning Machine Learning for Humans