资 源 简 介
大家好，我是Mac Jiang,今天和搭建分享的是台湾大学机器学习基石（Machine Learning Foundations）的个人笔记。个人觉得这门课是一门非常好的机器学习入门课程，值得初学者学习！这份笔记是本人一笔一划手写，扫描后上传了，也算是一个月的心血，希望我的工作能够给大家带来一些学习上的帮助。Week 1: The leaming ProblemDateP1. What is Machine Learning?m:画过观架( observation)款悍技能〔)cb→圆>skdem0过数(d枝能(s)t→DML→5m Prove:增进某种表现 Performance measurePer fornoMeasureML: an alternative rute to butt aomplacoded systems- When human connot prinn the system manually (navigating on Mars)-When human nnot define the solution esaily(speech/vgual recgnition)-When needing raBid deasions that huniang cannot do (high freguency trading?When neeling to be user-orrented in a massive sale(Consumer-targetd mopketig)dataim proved①有在其些日术Pattern RcRRMLerformance机购在完不知邮们度珠meusure的隐藏规剧料灿2.Appliaation of Machine learniO food: data Twitter data Wordsspill tell food Poisoning Cike lines of resturant properlye clothig. data: scale fiqures +client surveysskill: give good fashion recommendation to Clients3 Hosing, data, characteristic of buidings and their energy loadkill:predict energy lad of other buiding closely9 Trans potation: data: Some traffic sign images and meaningsShill= recognize troffic Sions acuratelyO Education: data. Studerts, records onizes ona math tutoring systemskilL: Predict whether a stdert can give a Comet answer to anther question⑥ ntertainment:da: ho w many users have hated some movies象社解料系子荐你统stiu Predict how a user Would rate an uNatPagc3. Com Ponent of Machine lemn输入:x∈x出:y∈Y睛数(9Mrtm)tf:y→丫〔想下的)抛规律台道数据如 raining examples:D=,,()…(xmhyes分sk:9x→y〔学到的程制的孤)辆一M→9Algur+nH( hy Pothesis Set)色色妇的成坏的Pt8,9∈H,从种中最的即9Leaming model= A and He hypothesis set4.Machine learning and other FiMachine (earring, B do值到约练于B数于的设3CMLatMn鸡:eg如 to find property that15e西不哦啥CDM)若立越西的为9R西事无大大区刷若栖与9关,PM可帮助MLArdt9g让电座有很瞰明的表视(下开)CA工)机学展现A工铝能的方法statistics(计利用瓷料爆到推龙,从数学角出发纯计晨钯机罟孑的方法第2讲: Learn to Answera人阳0 n Hypothesis set(假设集)Xxx)「y=(,许答 Wii threshold飞岁=+,讲卷Wx∠thr0BalePace3. Guarantee of PLAnear Separability(线性动):有在解线便对D的分和D展线性可分的■如果D是线性可分的台有在完全的使易=5(X)●证明线性可集D用PA反最危会侍下来個假设谢即鬼的,使h=gCM):有‰ t Wg Xnu)7 min yawl>,0()M更价当找到锘误,有hxm)→htMa≤o(3)通到销误点曼:Ww=+hnxa.每次纠正,w都在大出于对样本完美分割故h四x0,有: yntt)M Xnt)z min yM70当贴(Xm,如)进正:W=B4(+(e=W下+h时h即次销纠正,W都在增大一杜出一甘钟知m+2=+时2+2h←钟t#2b当愚剩得点后正为(X如),有9(0,即1和)≤Dtynut)nt=WIlt 2ynlt)Wt Xn)+(/ ynd Xn(b)-次Wt的坤长不可能无限大,限制。小侧M的增长,c游wt=w(+%m)xm+)ltmn为…27hMnW2=u+h1)42-w2+2lh:ym+x+t)14W会来松近((如tWT的龍Gs6但无限增大〔),散算法最金停下来DateP定义:R2=m刷2,P=米两,PA要来几次会倍下来如mn树x下fufil maIlX llT≤,最险时4. More about PLAPA优点宝现简单,速度快,任维度以下都可以行PA做点:①假D是线性司怕的一但们不D是线性胎,若不分PA会侮止⊙即使知道W在线惟印)但了知女会停下一P=m有2脏不知预久敬我们不道PA是子会角列使钟通公停,也不口多久才金停下②若D不艮钱性可分的么办?°/dyPA:Pet/gtbm(金心草法)--对子非线性司分D始化 Pocket weights0for t=O,①号找Wt首氓点CXnt)a⊙正错:W←-W+nu):xn③比W的错误率更小把更为直到尺的的送代次数 patio当知D是线性可分,则PA(垂度快),P邮历科本度慢连知D艰线性可分网用 Pocket第3#:下yP吁migl.Learning with Diferent Output SPace Ywto binary classficationn:y={-,+3· credit approve dis approvea email spam/nom--spamPatient Sick /not sickad Profitable/ds pofiubleanswer Correct I in Correct (EpD Cup Zolo)-()Multi class ClassificationCoin Reag nition problem y= fIc, SC, oC, 25c3=y=[L,2,--k3W比 gital classfication)· Picture(ore, strawberys emails(spam, Primary, SoCial Promotion, update)3)Regression: Y=RE Y= Clower, upperICR,Stock pmce predictionHouse Price Predictiontemperature Prediction) ructured learn的(绩柯t)复丧花式出Poid→ Protein foding(的结枸)Speeck dota→ speeck Prase tree2. Learning with Afferent Data lable yo, abelY supervised Learn iralt yn有鉴督学习:样库有本准宋y,CXn)组成样本元战督学:样本尺有X没有y,{xn组样库没有合造的分标准,谭作标准clustering:n3→ Cluster (x)6对立章自动桶其动融, density esti mation (oz*at)e9、对友通事故荡鬼的析Ot(t址tn(释常检测)門、信用卡湿刚检测Date(3) Semi-supervisedSome yh半监武学习。样库3准嘲,-没始(数face images with few labeled face identifier Gfoce book)mediclen data with a few (abded =medicine efect preditor原:全都进训练样本很嵌因抑很长酬只能标记小部你(4) rein farce ment Learning(h虽学)-量回报值为由稻增还学:叫4小为td,假假难狗h犹个确梅出,但告忙受错的设回报教若神定行得纬果网赵正回接丘若定进纤支售钱果则给自板,找剧摄最路在广和:(6 stomer, ad choice, click earning)纬广看用户是否击机器人表:对个动作台峰定回拔3. Learning with Different protoca子→ChhYcn batch learn底批习:把料全都给机,性学宛,9不化2) on Line learnioe:二样本样本相每收更→9会随索户味慢慢实毛PLA可以很宿易应用于线上学刁算法巾reinforce learning-都是线上成(岁地给回权笔一笔来)c3)active (earningaGe(amg,9 testion ask,相器自想皮发视新的M,f是升么对应从是什么用在取得@b很的场合,电的间我新画P就好了DatePage4 learning with Different Input space x(1) Raw Features原始着据:数锯不经任伺加工,接喂结机有1把像像泰为汉刷入() Abstract Features象据:没有物键义,对处理抽绿特像难曲妮定用对同歌曲评价,预测嚼的分,轴入为用取曲的分抽(3)Concrete Features且数据特:与物俚/际帱征关密切不发放信用卡,用户信息为泉体的征
