AI2013S_Chap05-Linear_Models_for_Classification英文文献资料.pdfVIP

Artificial Intelligence Linear Models for Classification Donghui Wang AI Institute@ZJU 2013 Spring Contents • Basic Concepts • Discriminant Functions • Probabilistic Generative Models • Probabilistic Discriminative Models • The Laplace Approximation • Bayesian Logistic Regression References: 1. Bishop. “Pattern Recognition and Machine Learning”, Chapter 4. 2006. Basic Concepts Linearly separable • Decision regions: – Input space is divided into several regions • Decision boundaries (surfaces): – Under linear models, it’s a linear function of the input vector x – (D-1)-dimensional hyper-plane within the D-dimensional input space • Data sets whose classes can be separated exactly by linear decision surfaces are said to be linear separable. T Representation of Class Labels • Two classes (K=2): – Target variable t ∈{0,1}, t=1 represents class C , else class C 1 2 • K-classes (K2): T W : weight vector w : bias/threshold – 1-of-K coding scheme: t = (0, 1, 0, 0, 0) 0 • Predict discrete class labels: T – Linear model prediction (linear discriminant function): y(x) = w x + w0 – Nonlinear function f (.) : R→(0,1) – Generalized linear models: f (·) : activation function f -1(·) : link function – Decisi