Neglected machine learning ideas

machine-learningonline-learningreinforcement-learningfeature-engineeringsurvey

Abstraction: Survey of underrepresented machine learning techniques missing from textbooks

Key points:

Online learning (sequential data exposure) is vastly undertreated in books; Vowpal Wabbit is a key tool; Cesa-Bianchi and Lugosi book covers theory
Reinforcement learning tied to control/robotics literature; TD-learning behind Backgammon success; Sutton and Barto insufficient for non-robotics use
Compression-based sequence prediction (LZW, Trie, Shannon) related to online learning; no unified ML text covers it
Conformal prediction provides well-motivated error bars on ML predictions; mostly work of Vovk/Shafer group
Feature engineering is crucial to ML success but has no dedicated textbook; SIFT, HOG, GIST, Hough transform important in vision
Unsupervised/semi-supervised learning (Self-Organizing Maps, semi-supervised for recommendation engines) largely ignored in mainstream books

Connections: Vowpal Wabbit · Online Learning · Reinforcement Learning · Feature Engineering · Unsupervised Learning · Conformal Prediction