机器学习前导---IBM讲座笔记

1. Machinelearning Application:

Recomendation

Compute Vision

Diriveless Car

Web Search

Speech Recagnition

Question Answering

Game Player

Smart Healthcare

Supervised learningand Unsupervised learning

eg:

classification,regression clustreing,association analysis

Machine learningtasks(problem types)

supervisedlearning=predict a target y from input x

(and seml-supervisedlearning)y represents a catrgory or “class”

=>classification

binary:y belongs to{-1,+1} or y belongs to (0,1)

multickass:y belongsto {1,m} or y belongs to {0,m-1}

y is a real-valuenumber

=>regression ybelongs to R or y belongs to R^m

Above is predictivemodels

Below isDesctrisptive models

Unsupervisedlearning :no explicit prediction target y

model the probablitydistribution of x

=>densityestimation

discover underlyingstructure in data

=>clustering

=>demensionalityreduction

=>(unsupervised )representation learning

trainging:we learn apredictive function f by optimizing it so that it predicts well onthe traing set

Use forprediction:we can then use in new inputs that were not part of thetraing set

关键词：generalize:泛化

the goal of learningis not to learn perfectly (memorizing) the traing set.What'simportant is the ability for the predictor to generide well on new

problem dimensions:

number of explits

input demensionditynumber if input features

characterizing eachexample

targetdimensionality ex:number of classes

under-fitting andoverfitting

关键词：fitting:拟合

performance ontraining set is not a good estimate of generalization

support vectormachine(SVM支持向量机）

LogisticRegression(逻辑斯蒂回归）

SVM：

A powerful methodfor classification

Bettergeneralization

key ideas

--use kernelfunctionto transform low dimensional training sample to higherdimensions

--use quadraticprogramming (QP二次规划）to find the best classifier

support vectors arethose dataprints that the margin pushes up against

the soft margin SVMis equivalent to applying a hinge loss

empirical loss +regularization 损失函数与正则化