UVA CS 6316 / CS 4501 – Fall 2015 : Machine Learning ... · Machine Learning Lecture 1: IntroducBon Dr. Yanjun Qi University of Virginia Department of ... Machine Learning 9/1/16

UVACS6316/CS4501–Fall2015:

MachineLearning

Lecture1:IntroducBon

Dr.YanjunQi

UniversityofVirginiaDepartmentof

ComputerScience

9/1/16

Dr.YanjunQi/UVACS6316-4501/f16

1

Welcome

•  CS6316/4501MachineLearning– MoWe3:30pm-4:45pm,– OlssonHall120

•  hOp://www.cs.virginia.edu/yanjun/teach/2016f

•  YourUVAcollab:Course6316-4501page

9/1/16


2

Today

q CourseLogisBcsq Mybackgroundq Basicsandroughcontentplanq ApplicaUonandHistory

9/1/16


3

CourseStaff

•  Instructor:Prof.YanjunQi– QI:/chee/– Youcancallme“professor”,“professorJane”,“professorQi”;

•  TAofficehours:Mon&Wed5:30pmpm-6:30pm@Rice504

•  Myofficehours:Mon5pm-6pm@Rice503

9/1/16


4

CourseLogisUcs

•  Courseemaillisthasbeensetup.Youshouldhavereceivedemailsalready!

•  Policy,thegradewillbecalculatedasfollows:– Assignments(55%,Sixtotal,each~9%)– Quizzes/ExamSamplePracUces(5%)– Midtermexam(20%)– Finalexam(20%)

9/1/16


5

CourseLogisUcs•  Midterm:Oct,75minsinclass•  Final:Dec,75minsinclass

•  Sixassignments(each9%to10%)– Threeextensiondayspolicy(checkcoursewebsite)

•  In-classquizzes/ExamsamplepracUce(total5%)

9/1/16


6

CourseLogisUcs•  Policy,

– HomeworkshouldbesubmiOedelectronicallythroughUVaCollab

– Homeworkshouldbefinishedindividually– Dueatmidnightontheduedate

–  Inordertopassthecourse,theaverageofyourmidtermandfinalmustalsobe"pass".

9/1/16


7

LateHomeworkPolicy

•  EachstudenthasthreeextensiondaystobeusedathisorherowndiscreUonthroughouttheenUrecourse.Yourgradeswouldbediscountedby15%perdaywhenyouusethese3latedays.Youcouldusethe3daysinwhatevercombinaUonyoulike.Forexample,all3dayson1assignment(foramaximumgradeof55%)or1eachdayover3assignments(foramaximumgradeof85%oneach).Aperyou'veusedall3days,youcannotgetcreditforanythingturnedinlate.

9/1/16


8

CourseLogisUcs

•  Textbooksforthisclassis:– NONE

•  Myslides–ifitisnotmenBonedinmyslides,itisnotanofficialtopicofthecourse

9/1/16


9

CourseLogisUcs

•  BackgroundNeeded– Calculus,Basiclinearalgebra,BasicprobabilityandBasicAlgorithm

– StaUsUcsisrecommended.– Studentsshouldalreadyhavegoodprogrammingskills,i.e.pythonisrequiredforallprogrammingassignments

– Wewillreview“linearalgebra”and“probability”inclass

9/1/16


10

Today

q CourseLogisUcsq Mybackground

q Basicsandroughcontentplanq ApplicaUonandHistory

9/1/16


11

AboutMe•  EducaUon:

– PhDfromSchoolofComputerScience,CarnegieMellonUniversity(@PiOsburgh,PA)in2008

– BSfromDepartmentofComputerScience,TsinghuaUniv.(@Beijing,China)

•  MyaccentPATTERN:/l/,/n/,/ou/,/m/•  Researchinterests:

– MachineLearning,BiomedicalapplicaBons

9/1/16


12

AboutMe

•  FiveYears’ofIndustryResearchLabinthepast:–  2008summer–2013summer,ResearchScienBstinITindustry(MachineLearningDepartment,NECLabsAmerica@Princeton,NJ)

–  2013Fall–Present,AssistantProfessor,ComputerScience,UVA

9/1/16

Industry + Academia


13

Today

q CourseLogisUcsq Mybackgroundq BasicsandRoughcontentplanq ApplicaUonandHistory

9/1/16


14

9/1/16


15

•  Biomedicine –  Patient records, brain imaging, MRI & CT scans, … –  Genomic sequences, bio-structure, drug effect info, …

•  Science

–  Historical documents, scanned books, databases from astronomy, environmental data, climate records, …

•  Social media

–  Social interactions data, twitter, facebook records, online reviews, …

•  Business –  Stock market transactions, corporate sales, airline traffic, …

•  Entertainment –  Internet images, Hollywood movies, music audio files, …

OUR DATA-RICH WORLD

9/1/16


16

•  Data capturing (sensor, smart devices, medical instruments, et al.)

•  Data transmission •  Data storage •  Data management •  High performance data processing •  Data visualization •  Data security & privacy (e.g. multiple

individuals) •  …… •  Data analytics

¢ How can we analyze this big data wealth ? ¢ E.g. Machine learning and data mining

BIG DATA CHALLENGES

this course

e.g.HCI

e.g.cloudcompuUng

BASICS OF MACHINE LEARNING

•  �The goal of machine learning is to build computer systems that can learn and adapt from their experience.� – Tom Dietterich

•  �Experience��in the form of available data examples (also called as instances, samples)

•  Available examples are described with properties (data points in feature space X)

9/1/16


17

9/1/16


18

e.g. SUPERVISED LEARNING

•  Find function to map input space X to output space Y

•  So that the difference between y and f(x)

of each example x is small.

Ibelievethatthisbookisnotatallhelpfulsinceitdoesnotexplainthoroughlythematerial.itjustprovidesthereaderwithtablesandcalculaUonsthatsomeUmesarenoteasilyunderstood…

x

y-1

InputX:e.g.apieceofEnglishtext

OutputY:{1/Yes,-1/No}e.g.IsthisaposiUveproductreview?

e.g.

e.g. SUPERVISED Linear Binary Classifier

f x y

f(x,w,b) = sign(w x + b)

wx+b<0

CourtesyslidefromProf.AndrewMoore’stutorial

?

?

wx+b>0

denotes +1 point

denotes -1 point

denotes future points

?

9/1/16


19

9/1/16


20

•  Training (i.e. learning parameters w,b ) –  Training set includes

•  available examples x1,…,xL •  available corresponding labels y1,…,yL

– Find (w,b) by minimizing loss (i.e. difference between y and f(x) on available examples in training set)

(W, b) = argminW, b

Basic Concepts

•  Testing (i.e. evaluating performance on �future��points) –  Difference between true y? and the predicted f(x?) on a

set of testing examples (i.e. testing set)

–  Key: example x? not in the training set

•  Generalisation:learnfuncUon/hypothesisfrompastdatainorderto“explain”,“predict”,“model”or“control”newdataexamples

9/1/16

Basic Concepts Dr.YanjunQi/UVACS6316-4501/f16

21

9/1/16


22

•  Loss function –  e.g. hinge loss for binary

classification task

–  e.g. pairwise ranking loss for

ranking task (i.e. ordering examples by preference)

•  Regularization –  E.g. additional information added on loss function to control model

Basic Concepts

TYPICAL MACHINE LEARNING SYSTEM

9/1/16

Low-level sensing

Pre-processing

Feature Extract

Feature Select

Inference, Prediction, Recognition

Label Collection


23

Evaluation

“BigData”ChallengesforMachineLearning

9/1/16

ü Largesizeofsamplesü Highdimensionalfeatures

Not the focus, will be covered in advanced-level course


24

Large-Scale Machine Learning: SIZE MATTERS

25

9/1/16

•  One thousand data instances

•  One million data instances

•  One billion data instances

• One trillion data instances

Thosearenotdifferentnumbers,thosearedifferentmindsets!!!


BIG DATA CHALLENGES FOR MACHINE LEARNING

9/1/16

ThevariaUonsofbothX(feature,representaUon)andY(labels)arecomplex!

Most of this

course ü ComplexityofXü ComplexityofY


26

TYPICAL MACHINE LEARNING SYSTEM

9/1/16

Low-level sensing

Pre-processing

Feature Extract

Feature Select

Inference, Prediction, Recognition

Label Collection

Data Complexity of X

Data Complexity

of Y


27

Evaluation

UNSUPERVISED LEARNING :

[ COMPLEXITY OF Y ]

•  No labels are provided (e.g. No Y provided)

•  Find patterns from unlabeled data, e.g. clustering

9/1/16

e.g.clustering=>tofind�natural�groupingofinstancesgivenun-labeleddata


28

STRUCTURAL OUTPUT LEARNING : [ COMPLEXITY OF Y ]

•  Many prediction tasks involve output labels having structured correlations or constraints among instances

9/1/16

Manymorepossiblestructuresbetweeny_i,e.g.spaUal,temporal,relaUonal…

Thedogchasedthecat

APAFSVSPASGACGPECA…

TreeSequence GridStructured Dependency between Examples

Input

Output

CCEEEEECCCCCHHHCCC…


29

Original Space Feature Space

�

�

�

STRUCTURAL INPUT : Kernel Methods [ COMPLEXITY OF X ]

Vectorvs.RelaUonaldata

e.g.Graphs,Sequences,3Dstructures,

9/1/16


30

MORE RECENT: FEATURE LEARNING [ COMPLEXITY OF X ]

Deep Learning Supervised Embedding

9/1/16

Layer-wise Pretraining


31

DEEP LEARNING / FEATURE LEARNING : [ COMPLEXITY OF X ]

Feature Engineering ü  Most critical for accuracy ü  Account for most of the computation for testing ü  Most time-consuming in development cycle ü  Often hand-craft and task dependent in practice

Feature Learning ü  Easily adaptable to new similar tasks ü  Layerwise representation ü  Layer-by-layer unsupervised training ü  Layer-by-layer supervised training 329/1/16


339/1/16


CourseContentPlanèFivemajorsecUonsofthiscourse

q Regression(supervised)q ClassificaUon(supervised)q Unsupervisedmodelsq Learningtheoryq Graphicalmodels

9/1/16 34


9/1/16 35

hOp://scikit-learn.org/Dr.YanjunQi/UVACS6316-4501/f16

Scikit-learnalgorithmcheat-sheet

9/1/16 36

hOp://scikit-learn.org/stable/tutorial/machine_learning_map/Dr.YanjunQi/UVACS6316-4501/f16

9/1/16 37

Scikit-learn:Regression

LinearmodelfiOedbyminimizingaregularizedempiricallosswithSGD


LinearRegression+VariaUons

Scikit-learn:ClassificaUon

9/1/16 38

Linearclassifiers(SVM,logisUcregression…)withSGDtraining.

approximatetheexplicitfeaturemappingsthatcorrespondtocertainkernelsTocombinethe

predicUonsofseveralbaseesUmatorsbuiltwithagivenlearningalgorithminordertoimprovegeneralizability/robustnessoverasingleesUmator.(1)averaging/bagging(2)boosUng


9/1/16 39

BasicPCA

Bayes-NetHMM

Kmeans+GMM

UnsupervisedModelsDr.YanjunQi/UVACS6316-4501/f16

Summary

•  Thisisnotacourseaboutlearningtousetoolbox

•  Wefocusonlearningprinciples,mathemaUcalformulaUon,algorithmdesignandlearningtheory.

9/1/16


40

Today

q CourseLogisUcsq Mybackgroundq Basicsandroughcontentplanq ApplicaBonandHistory

9/1/16


41

Whatcanwedowiththedatawealth?èREAL-WORLDIMPACT

§ Businessefficiencies§  ScienUficbreakthroughs§  Improvequality-of-life:§  healthcare,§  energysaving/generaUon,§  environmentaldisasters,§  nursinghome,§  transportaUon,§  …

9/1/16

MedicalImages

GenomicData

TransportaUonData

BraincomputerinteracUon(BCI)

Devicesensordata


42

When to use Machine Learning (Adaptto/learnfromdata) ?

–  1. Extract knowledge from data –  Relationships and correlations can be hidden within large

amounts of data –  The amount of knowledge available about certain tasks is

simply too large for explicit encoding (e.g. rules) by humans

–  2. Learn tasks that are difficult to formalise –  Hard to be defined well, except by examples, e.g., face

recognition

–  3. Create software that improves over time –  New knowledge is constantly being discovered. –  Rule or human encoding-based system is difficult to

continuously re-design �by hand�.

9/1/16


43

MACHINE LEARNING IS CHANGING THE WORLD

9/1/16 Manymore!


44

MACHINE LEARNING IN COMPUTER SCIENCE

•  Machine learning is already the preferred approach for

–  Speech recognition, natural language processing –  Computer vision –  Medical outcome analysis –  Robot control …

•  Why growing ? –  Improved machine learning algorithms –  Increased data capture, new sensors, networking –  Systems/Software too complex to control manually –  Demand to self-customization for user, environment, ….

9/1/16


45

46

RELATED DISCIPLINES

•  Artificial Intelligence •  Data Mining •  Probability and Statistics •  Information theory •  Numerical optimization •  Computational complexity theory •  Control theory (adaptive) •  Psychology (developmental, cognitive) •  Neurobiology •  Linguistics •  Philosophy

9/1/16


WhatarethegoalsofAIresearch?

ArUfactsthatACTlikeHUMANS

ArUfactsthatTHINKlikeHUMANS

ArUfactsthatTHINKRATIONALLY

ArUfactsthatACTRATIONALLY

47From:M.A.Papalaskar

9/1/16


Howcanwebuildmoreintelligentcomputer/machine?

•  Ableto–  perceivetheworld–  understandtheworld

•  Thisneeds–  BasicspeechcapabiliUes–  BasicvisioncapabiliUes–  Language/semanUcunderstanding– Userbehavior/emoUonunderstanding–  Abletothink??

9/1/16 48


9/1/16 49


toservehumanbeings,and

fluentin"oversixmillionformsofcommunicaUon"

R2-D2andC-3PO

@StarWars–1977



9/1/16 50

Jeopardy Game è Requires a Broad Knowledge Base

IBM Watson è an artificial intelligence computer system capable of answering questions posed in natural language developed in IBM's DeepQA project.



51

Apple Siri / Amazon Echo è an intelligent personal assistant and knowledge navigator

9/1/16


Howcanwebuildmoreintelligentcomputer/machine?: Objective Recognition / Image Labeling

Deep Convolution Neural Network (CNN) won (as Best systems) on �very large-scale��ImageNet competition 2012 / 2013 / 2014

89%, 2013

529/1/16


ImageNet:animagedatabaseorganizedaccordingtotheWordNetLSVRC:LargeScaleVisualRecogniUonChallengebasedonImageNet.

93%, 2014 [ training on 1.2 million images [X] vs. 1000 different word labels [Y] ]

Howcanwebuildmoreintelligentcomputer/machine?: Objective Recognition / Image Labeling

9/1/16


53

-  2013,GoogleAcquiredDeepNeuralNetworksCompanyheadedbyUtoronto“DeepLearning”ProfessorHinton

-  2013,FacebookBuiltNewArUficialIntelligenceLabheadedbyNYU“DeepLearning”ProfessorLeCun

-  2016,Google'sDeepMinddefeatslegendaryGoplayerLeeSe-dolinhistoricvictory

Detour:plannedprogrammingassignments

•  HW:SemanUclanguageunderstanding(senUmentclassificaUononmoviereviewtext)

•  HW:VisualobjectrecogniUon(labelingimagesabouthandwriOendigits)

•  HW:AudiospeechrecogniUon(unsupervisedlearningbasedspeechrecogniUontask)

9/1/16 54


TodayRecap

q CourseLogisUcsq Mybackgroundq Basicsandroughcontentplanq ApplicaUonandHistory

9/1/16


55

Task

Next lesson: Machine Learning in a Nutshell

Representation

Score Function

Search/Optimization

Models, Parameters

9/1/16 56

MLgrewoutofworkinAIOp+mizeaperformancecriterionusingexampledataorpastexperience,Aimingtogeneralizetounseendata


Next lesson: Review of linear algebra and basic calculus

References

q Prof.AndrewMoore’stutorialsq Prof.RaymondJ.Mooney’sslidesq Prof.AlexanderGray’sslidesq Prof.EricXing’sslidesq hOp://scikit-learn.org/q HasUe,Trevor,etal.TheelementsofstaUsUcallearning.Vol.2.No.1.NewYork:Springer,2009.

q Prof.M.A.Papalaskar’sslides

9/1/16


57

UVA CS 6316 / CS 4501 – Fall 2015 : Machine Learning ... · Machine Learning Lecture 1: IntroducBon Dr. Yanjun Qi University of Virginia Department of ... Machine Learning 9/1/16

Documents