Wrap UpWrap Up

Introduction toIntroduction toArtificial IntelligenceArtificial Intelligence

COS302COS302

Michael L. LittmanMichael L. Littman

Fall 2001Fall 2001

AdministrationAdministration

Laptop four is busted.Laptop four is busted.

Course evaluations.Course evaluations.

PlanPlan

LSI and autoassociationLSI and autoassociation

Topic listTopic list

Second half ThemesSecond half Themes

Wrap upWrap up

Passage AutoassociatorPassage Autoassociator

xx11

hh11

xx22 xx33 xx44

hh22

U (nxk)U (nxk)

UUT T (kxn)(kxn)

xx11 xx22 xx33 xx44

word featuresword features

capture “gist” so it capture “gist” so it can be can be reconstructedreconstructed

We’ll Make Computers We’ll Make Computers Brighter Brighter (sorry Billy Joel)(sorry Billy Joel)

Gang and Kedar relaxationGang and Kedar relaxation color graphs by propagationcolor graphs by propagation

Neighbor goal state Davis Putnam Neighbor goal state Davis Putnam change to CNFchange to CNF

CSP & satisfactionCSP & satisfaction forward checking giving tractionforward checking giving traction

BFS and DFS andBFS and DFS andA* is the bestA* is the best

Verse 2Verse 2

Local search save timeLocal search save time optimize hill climboptimize hill climb

Nim Sharks GSATNim Sharks GSAT chess chess evaluation hacksevaluation hacks

Alpha beta game treeAlpha beta game tree prune prune the branch and you’ll seethe branch and you’ll see

Fitness function reproduction phase Fitness function reproduction phase transition min-maxtransition min-max

ChorusChorus

We’ll make computers brighterWe’ll make computers brighter

With their “and” and “or”ingWith their “and” and “or”ing

They’ll beat Alan Turing.They’ll beat Alan Turing.

We’ll make computers brighterWe’ll make computers brighter

Smarter than a porpoiseSmarter than a porpoise

Train it with a corpusTrain it with a corpus

Verse 3Verse 3

Markov model Chinese RoomMarkov model Chinese RoomNim-Rand in a matrix formNim-Rand in a matrix form

Labeled data, missing dataLabeled data, missing datalogic of Boolelogic of Boole

Discount rank list Discount rank list Altavista Altavista WordnetWordnet

IR bag-of-words IR bag-of-words flip around with Bayes Ruleflip around with Bayes Rule

Verse 4Verse 4

Rap song n-queenRap song n-queenSusan swallowed something greenSusan swallowed something green

Language model classifyLanguage model classify take take a max and multiplya max and multiply

Zipf’s law trigramsZipf’s law trigramssmooth it out with bigramssmooth it out with bigrams

Naïve Bayes winning playsNaïve Bayes winning plays write write another programanother program

ChorusChorus

We’ll make computers brighterWe’ll make computers brighter

With their “and” and “or”ingWith their “and” and “or”ing

They’ll beat Alan Turing.They’ll beat Alan Turing.

We’ll make computers brighterWe’ll make computers brighter

Smarter than a porpoiseSmarter than a porpoise

Train it with a corpusTrain it with a corpus

Verse 5Verse 5

HMM part-of-speechHMM part-of-speechthey will learn if you will teachthey will learn if you will teach

Baum-Welch if you pleaseBaum-Welch if you please Rush Rush Hour and AnalogiesHour and Analogies

Weighted coins network lagsWeighted coins network lagsViterbi finds most likely tagsViterbi finds most likely tags

Start running watch for tiesStart running watch for tiesexpectation maximizeexpectation maximize

Verse 6Verse 6

Learning function featureLearning function featuretry learning who’s a girltry learning who’s a girl

overfit PAC boundsoverfit PAC boundsbuild your own decision treebuild your own decision tree

Too big memorize Too big memorize cross validate to generalizecross validate to generalize

supervising pure leafsupervising pure leaf fit fit the function to a ‘t’the function to a ‘t’

ChorusChorus

We’ll make computers brighterWe’ll make computers brighter

With their “and” and “or”ingWith their “and” and “or”ing

They’ll beat Alan Turing.They’ll beat Alan Turing.

We’ll make computers brighterWe’ll make computers brighter

Smarter than a porpoiseSmarter than a porpoise

Train it with a corpusTrain it with a corpus

Verse 7Verse 7

Delta rule backpropDelta rule backprophidden unit don’t stophidden unit don’t stop

Learn pairs sum squaresLearn pairs sum squareslinear perceptronlinear perceptron

Error surface is quite bentError surface is quite bentjust descend the gradient and-or just descend the gradient and-or sigmoidsigmoidnonlinear regressionnonlinear regression

Bridge/ChorusBridge/Chorus

XOR learning rateXOR learning ratesquash the sum and integratesquash the sum and integrate

Neural net split the setNeural net split the setHow much better can it get?How much better can it get?

ChorusChorus: We’ll make computers brighter: We’ll make computers brighterWith their “and” and “or”ingWith their “and” and “or”ingThey’ll beat Alan Turing.They’ll beat Alan Turing.We’ll make computers brighterWe’ll make computers brighterSmarter than a porpoiseSmarter than a porpoiseTrain it with a corpusTrain it with a corpus

Verse 8Verse 8

LSI max and minLSI max and minlocal search is back againlocal search is back again

TOEFL pick bestTOEFL pick best essay essay grade pass the testgrade pass the test

Matrix Plato SVDMatrix Plato SVD vectors vectors all in high Dall in high D

All the points are in the spaceAll the points are in the spacedon’t you make an Eigenface!don’t you make an Eigenface!

Verse 9Verse 9

Segmentation Saturn spinSegmentation Saturn spinProbabilities a winProbabilities a win

Q and P iterativelyQ and P iterativelyBayes net CPTBayes net CPT

Autopilots in the skyAutopilots in the skyPlay chess masters for a tiePlay chess masters for a tieAlan Alda robots cryAlan Alda robots cryAll of this is in AI!All of this is in AI!

ChorusChorus

We’ll make computers brighterWe’ll make computers brighter

With their “and” and “or”ingWith their “and” and “or”ing

They’ll beat Alan Turing.They’ll beat Alan Turing.

We’ll make computers brighterWe’ll make computers brighter

Smarter than a porpoiseSmarter than a porpoise

Train it with a corpusTrain it with a corpus

Final ChorusFinal Chorus

We’ll make computers brighterWe’ll make computers brighter

Hope you had some funHope you had some fun

Because the class is done, is done, is Because the class is done, is done, is done…done…

Probability ModelsProbability Models

Compare and contrastCompare and contrast• unigramunigram• bigrambigram• trigramtrigram• HMMHMM• belief netbelief net• Naïve BayesNaïve Bayes

Classification AlgsClassification Algs

Compare and contrastCompare and contrast• Naïve BayesNaïve Bayes• decision treedecision tree• perceptronperceptron• multilayer neural netmultilayer neural net

Missing DataMissing Data

Compare and contrastCompare and contrast• EMEM• SVDSVD• gradient descentgradient descent

Dynamic ProgrammingDynamic Programming

Compare and contrastCompare and contrast• ViterbiViterbi• forward-backwardforward-backward• minimaxminimax• Markov chains (value iteration)Markov chains (value iteration)• segmentationsegmentation

Big Picture: This is AI?Big Picture: This is AI?

We learned a bunch of programming We learned a bunch of programming tricks: people still do the hard work.tricks: people still do the hard work.

• Yeah, that’s true.Yeah, that’s true.• It’s not such a bad thing: building more It’s not such a bad thing: building more

flexible software (search, learning).flexible software (search, learning).• We’re working towards more We’re working towards more

independent artifacts; not there yetindependent artifacts; not there yet

Thank You!Thank You!

Thanks for your help!Thanks for your help!• Gang and KedarGang and Kedar• Lisa and the kidsLisa and the kids• AT&TAT&T• Andrew MooreAndrew Moore• Family Feud survey respondentsFamily Feud survey respondents• Peter StonePeter Stone• all of you!all of you!