Lecture 4: Learning Theory (cont.) and optimization … · Lecture 4: Learning Theory (cont.) and optimization (start) Sanjeev Arora Elad Hazan COS 402 –Machine Learning and Artificial

Lecture4:LearningTheory(cont.)andoptimization(start)

SanjeevArora EladHazan

COS402– MachineLearningand

ArtificialIntelligenceFall2016

Admin

• Exercise1– duetoday• Exercise2(implementation)nextTue,inclass• Enrolment…• Latepolicy(exercises)• Literature(free!)

Recap

Lastlecture:• ShiftfromAIbyintrospection(Naïvemethods)tostatistical/computational learningtheory

• Fundamentaltheoremofstatisticallearning• Samplecomplexity,overfitting,generalization

Agenda

• (Review)statistical&computationallearningtheoriesforlearningfromexamples

• (Review)Fundamentaltheoremofstatisticallearningforfinitehypothesisclasses

• Theroleofoptimizationinlearningfromexamples• Linearclassificationandtheperceptron• SVM andconvexrelaxations

Definition:learningfromexamplesw.r.t.hypothesisclassAlearningproblem:𝐿 = (𝑋,𝑌, 𝑐, ℓ,𝐻)• X=Domainofexamples(emails,pictures,documents,…)• Y=labelspace(forthistalk,binaryY={0,1})• D=distributionover(X,Y)(theworld)• Dataaccessmodel:learnercanobtaini.i.d samplesfromD• Concept=mapping𝑐:𝑋 ↦ 𝑌• Lossfunctionℓ: 𝑌, 𝑌 ↦ 𝑅,suchasℓ 𝑦/, 𝑦0 = 123425• H=classofhypothesis:𝐻 ⊆ {𝑋 ↦ 𝑌}• Goal:producehypothesish∈ 𝐻withlowgeneralizationerror

𝑒𝑟𝑟 ℎ = 𝐸 ?,2 ∼A[ℓ(ℎ 𝑥 , 𝑐 𝑥 )]

agnosticPAClearnability

Learningproblem𝐿 = (𝑋,𝑌,𝐻, 𝑐, ℓ)isagnosticallyPAC-learnable ifthereexistsalearningalgorithms.t.forevery𝛿, 𝜖 > 0,thereexistsm = 𝑓 𝜖, 𝛿,𝐻 < ∞,s.t. afterobservingSexamples,for 𝑆 = 𝑚,returnsahypothesisℎ ∈ 𝐻,suchthatwithprobabilityatleast

1 − 𝛿itholdsthat

𝑒𝑟𝑟 ℎ ≤ minZ∗∈\

𝑒𝑟𝑟 ℎ∗ + 𝜖

Equalszeroforrealizableconcepts

Themeaningoflearningfromexamples

Theorem:Everyrealizablelearningproblem𝐿 = (𝑋,𝑌,𝐻, 𝑐, ℓ)forfiniteH,isPAC-learnablewithsample

complexity𝑆 = 𝑂_`a\b_`a3c

d usingtheERMalgorithm.

Noamchomesky, June2011:“It'struethere'sbeenalotofworkontrying toapplystatisticalmodelstovariouslinguisticproblems.Ithinktherehavebeensomesuccesses,butalotoffailures.Thereisanotion ofsuccess...whichIthinkisnovelinthehistoryofscience.Itinterpretssuccessasapproximatingunanalyzeddata.”

Examples– statisticallearningtheorem

Theorem:Everyrealizablelearningproblem𝐿 = (𝑋,𝑌,𝐻, 𝑐, ℓ)forfiniteH,isPAC-learnablewithsample

complexity𝑆 = 𝑂_`a\b_`a3c

d usingtheERMalgorithm.

• Applefactory:Wt.ismeasuredingrams,100-400scale.Diameter:centimeters,3-20

• Spamclassificationusingdecisiontreesofsize20nodes• 200Kwords

Infinitehypothesisclasses

VC-dimension:correspondingto“effectivesize”ofhypothesisclass(infiniteorfinite)• Finiteclasses,𝑉𝐶𝑑𝑖𝑚 𝐻 = log𝐻• Axis-alignedrectanglesin𝑅j ,𝑉𝐶𝑑𝑖𝑚 𝐻 = 𝑂(𝑑)• Hyperplanesin𝑅j ,𝑉𝐶𝑑𝑖𝑚 𝐻 = 𝑑 + 1• Polygonsintheplane,𝑉𝐶𝑑𝑖𝑚 𝐻 = ∞

Fundamentaltheoremofstatisticallearning:Arealizablelearningproblem𝐿 = (𝑋, 𝑌, 𝐻, 𝑐, ℓ) isPAC-learnable ifanonlyifitsVCdimensionisfinite,

inwhichitislearnablewithsamplecomplexity𝑆 = 𝑂klmno(\)_`a3c

d usingtheERMalgorithm.

Overfitting??

samplecomplexity𝑆 = 𝑂_`a\b_`a3c

d

Itistight!

Reminder:classifyingfuelefficiency

• 40datapoints

• Goal:predictMPG

mpg cylinders displacement horsepower weight acceleration modelyear maker

good 4 low low low high 75to78 asiabad 6 medium medium medium medium 70to74 americabad 4 medium medium medium low 75to78 europebad 8 high high high low 70to74 americabad 6 medium medium medium medium 70to74 americabad 4 low medium low medium 70to74 asiabad 4 low medium low low 70to74 asiabad 8 high high high low 75to78 america: : : : : : : :: : : : : : : :: : : : : : : :bad 8 high high high low 70to74 americagood 8 high medium high high 79to83 americabad 8 high high high low 75to78 americagood 4 low low low low 79to83 americabad 6 medium medium medium high 75to78 americagood 4 medium low low low 79to83 americagood 4 low low medium high 79to83 americabad 8 high high high low 70to74 americagood 4 low medium low medium 75to78 europebad 5 medium medium medium medium 75to78 europe

X Y

The test set error is much worse than the training set error…

…why?

Figure fromMachineLearningandPatternRecognition,Bishop

𝑡 = sin 2𝜋𝑥 + 𝜖

𝑡 = sin 2𝜋𝑥 + 𝜖

Figure fromMachineLearningandPatternRecognition,Bishop

Figure fromMachineLearningandPatternRecognition,Bishop

Occam’srazor

WilliamofOccam c. 1287– 1347:controversialtheologian:“pluralityshouldnotbepositedwithoutnecessity”,i.e.“thesimplestexplanationisbest”

Theorem:Everyrealizablelearningproblem𝐿 = (𝑋,𝑌,𝐻, 𝑐, ℓ)forfiniteH,isPAC-learnablewithsample

complexity𝑆 = 𝑂_`a\b_`a3c

d usingtheERMalgorithm.

Otherhypothesisclasses?

1. Pythonprogramsof<=10000wordsH ≈ 10000/uuuu

à samplecomplexityis𝑂_`a\b_`a3c

d = 𝑂(vuwd ) - nottoobad!

2. Efficientalgorithm?3. (Haltingproblem…)

4. TheMAINissuewithPAClearningiscomputationalefficiency!5. Nexttopic:MOREHYPOTHESISCLASSESthatpermitefficientOPTIMIZATION

Booleanhypothesis

Monomialonabooleanfeaturevector:

𝑀 𝑥/, . . , 𝑥z = �̅�z ∧ 𝑥0 ∧ 𝑥/

(homeworkexercise…)

x1 x2 x3 x4 y

1 0 0 1 0

0 1 0 0 1

0 1 1 0 1

1 1 1 0 0

Linearclassifiers

Domain=vectorsoverEuclideanspaceRm

Hypothesisclass:allhyperplanesthatclassifyaccordingto:

ℎ 𝑥 = 𝑠𝑖𝑔𝑛(𝑤�𝑥 − 𝑏)

(weusuallyignoreb– thebias,itis0almostw.l.o.g.)

Empirically:theworldismanytimeslinearly-separable

Thestatisticsoflinearseparators

Hyperplanesinddimensionswithnormatmost1,andaccuracyof𝜖:

𝐻 ≈1𝜖

j

VCdimensionisd+1

è Samplecomplexity𝑂(jb_`a3c

d )

Findingthebestlinearclassifier:LinearClassification

TheERMalgorithmreducesto:

n vectorsind dimensions: x/,x0, … , x� ∈ 𝑅j

Labelsy/,𝑦0, … , 𝑦� ∈ −1,1 Findvectorwsuchthat:

∀𝑖. 𝑠𝑖𝑔𝑛 𝑤�𝑥� = 𝑦�

Assume: w ≤ 1, 𝑥� ≤ 1

𝜖= margin

The margin

Margin 𝜖 = max � �/

min�{𝑤�𝑥� ⋅ 𝑦�}

The Perceptron Algorithm[Rosenblatt 1957, Novikoff 1962, Minsky&Papert 1969]

The Perceptron Algorithm

Iteratively:1. Find vector 𝑥� for which sig𝑛 𝑤�𝑥� ≠ 𝑦�2. Add 𝑥� to w:

𝑤�b/ ← 𝑤� + 𝑦�𝑥�

The Perceptron Algorithm

Thm [Novikoff 1962]: for data with margin 𝜖, perceptron

returns separating hyperplane in /d5 iterations

Thm [Novikoff 1962]: converges in 1/ 𝜖0 iterationsProof:

Let w* be the optimal hyperplane, s.t. ∀𝑖, ynxn�w∗ ≥ 𝜖

1. Observation 1: w�b/� w∗ = w� + y�x� w∗ ≥ 𝑤��𝑤∗ + 𝜖

2. Observation 2: |w�b/

� |0 = |w��|0 + 𝑦�𝑥��𝑤� + 𝑦�𝑥� 0 ≤ |w�

�|0+1Thus:

1 ≥𝑤�|𝑤�|

⋅ w∗ ≥𝑡𝜖𝑡= 𝑡𝜖

And hence: t ≤ /d5

Perceptron:𝑤�b/ ← 𝑤� +𝑦�𝑥�

Noise?

ERMfornoisylinearseparators?

Givenasample𝑆 = 𝑥/,𝑦/ , … , 𝑥�,𝑦� ,findhyperplane(throughtheoriginw.l.o.g)suchthat:

𝑤 = arg min� �/

𝑖s. 𝑡. 𝑠𝑖𝑔𝑛(𝑤�𝑥� ≠ 𝑦� |

• NP-hard!

• à convexrelaxation+optimization!

Noise– minimizesumofweightedviolations

Soft-marginSVM(supportvectormachines)

Givenasample𝑆 = 𝑥/,𝑦/ , … , 𝑥�,𝑦� ,findhyperplane(throughtheoriginw.l.o.g)suchthat:

𝑤 = arg min� �/

{1𝑚�max{0,1 −𝑦�𝑤�𝑥�}

�}

• Efficientlysolvablebygreedyalgorithm– gradientdescent• Moregeneralmethodology:convexoptimization

Summary

PAC/Statisticallearningtheory:• Precisedefinitionoflearningfromexample

• Powerful&verygeneralmodel• Exactcharacterizationof#ofexamplestolearn(samplecomplexity)• Reductionfromlearningtooptimization• Arguedfinitehypothesisclassesarewonderful(Python)• Motivatedefficientoptimization• LinearclassificationandthePerceptron+analysis• SVMà convexoptimization(nexttime!)