无监督学习总结.pdf

A UNIFIED PERSPECTIVE ON SUPERVISED, UNSUPERVISED AND SEMI-SUPERVISED LEARNING Problem Types Problem Types Supervised learning Regression Given input descriptions x 2 Rn and target values y 2 Rk y y y y x x x x Infer a function f : X ! Y Usually k n Problem Types Supervised Unsupervised learning Regression Dimensionality reduction Given input descriptions x 2 Rn y y y y x x x x Simultaneously infer a function f : X ! Y and a guess at the missing labels Z Unsupervised regression usually called dimensionality reduction because k n Problem Types Supervised learning Classification Given inputs x 2 Rn and target values y 2 f1; :::; kg y y y y x x x x Infer a function f : X ! Y Usually k n Problem Types Supervised Unsupervised learning Classification Clustering Given inputs x 2 Rn y y y y x x x x Simultaneously infer a function f : X ! Y and a guess at the missing labels Z Unsupervised classification usually called clustering because the guessed classifications Z are emphasized over f Problem Types Semi-supervised learning Some training labels are given, but most are not y y y y y y x x x x x x Training Principles  Are unsupervised and supervised training problems really so unrelated?  Does supervised and unsupervised learning require distinct training principles, distinct motivations, distinct theory?  Maybe not Key Challenge Relating unsupervised to supervised learning 1. If both