Robust Algorithms for Machine Learning
summary
Argues that throwing away cardinal values often helps: replace values with ranking and counting or signs to resist outliers, use SciPy stats for Mann‑Whitney, Kruskal‑Wallis, Kolmogorov and feed robust features into Keras/keras backends. Gives two concrete results: an automated water‑bacteria detector is made fault tolerant and a stock test on ~1,800 trading days finds 18 vs 1 expected overlap (7 observed) between halves. Notes practical tooling: ActivePython with pandas and scipy.stats and that a neural net trained on raw signs only hits baseline accuracy.
AI-written summary · about
chapters (8)
- 0:00Robust algorithms intro
- 7:27Why robust algorithms matter
- 15:38Parametric limits with noise
- 21:20Non-parametric robust tests
- 29:07Practical robustness applications
- 35:05Deep learning bias caution
- 38:22Meta-overtraining pitfalls
- 40:48Q&A
related talks
30mMachine Learning Techniques for Class Imbalances & Adversaries
40mTools and Tricks from a Pragmatic Data Scientist
30mMachine Learning: Beyond the Tutorial
53mResponsible AI with SciPy
44mHandling predictive uncertainty in Machine Learning
42mMachine Learning Crash Course
42mExploring Pillars of Trustworthy AI: Robustness and Fairness
45m"Breaking black-box machine learning"