Practical concerns in predicting structures...Practical concerns in predicting ... – We want to convert some input to an output – Model the conditional distribution of the output

CS6961:StructuredPredictionSpring2017

Practicalconcernsinpredictingstructures

1

Sofar…

• Whatarestructures?– Agraph– Acollectionofpartsthatarelabeledjointly– Acollectionofinterconnecteddecisions

• Conditionalmodels– Wewanttoconvertsomeinputtoanoutput– Modeltheconditionaldistributionoftheoutput– Scoregroupsofinter-connectedvariables

• Algorithmsforlearning– Localvs.globallearning– Differentalgorithms

• Inferencealgorithms– Predictingthefinaloutput– Differentalgorithms,tradeoffs

2

Usingthetools:Practicalconcerns

• Whatarestructures?– Agraph– Acollectionofpartsthatarelabeledjointly– Acollectionofinterconnecteddecisions

• Conditionalmodels– Wewanttoconvertsomeinputtoanoutput– Modeltheconditionaldistributionoftheoutput– Scoregroupsofinter-connectedvariables

• Algorithmsforlearning– Localvs.globallearning– Differentalgorithms

• Inferencealgorithms– Predictingthefinaloutput– Differentalgorithms,tradeoffs

3

Thislecture:

• Wewanttosolveatask.• Fornow,fullysupervised• Manychoicesahead!

Whatisthegraph?

Thebestwaytolearn?

• Modelingourproblem?• Identifyingvariables?• Identifyinggroupsthatare

scoredtogether?(factors)• Whatarefeatures?

Whatinferencealgorithm?

DoingeverythinginprogramminglanguageX

Modelingyourproblem

• Understandtheproblem:Whatshouldyourprogramproduce?– Istheredata?Veryoften,theanswerisno.L

• Whatarethedecisions/randomvariablesthatconstitutetheoutput?

• Howdotheyinteract?Identifyingfactors/parts– Someinteractionsarenatural,somearespurious(specifictoyoursmallcollectionofdata)– Someinteractionsmakeinferenceimpossibleforcomputationalreasons– Whatarethefeaturerepresentations?

• Learning– Shouldeveryscoringfunctionbejointlylearned?– Perhaps,learnsub-sectionsindependentlyandputthemtogetherwithinferenceattheend– Whichlearningalgorithm?

• Inference– Whatalgorithm?Howexpensiveisit?– Exactorapproximate?

4

Example0:NamedEntityRecognition

Goal:Toidentifypersons,locationsandorganizationsintext

B-org O B-per I-per O O

Facebook CEOMarkZuckerberg announcednewO O O O O O B-loc I-loc

privacy featuresintheconferenceinSanFrancisco

5

ObservationsStates

B-org=StartoforganizationB-per=StartofpersonI-per=Inperson

B-loc =StartoflocationI-loc =InlocationO=Notanamedentity

Example0:NamedEntityRecognition

Goal:Toidentifypersons,locationsandorganizationsintext

B-org O B-per I-per O O

Facebook CEOMarkZuckerberg announcednewO O O O O O B-loc I-loc

privacy featuresintheconferenceinSanFrancisco

6

ObservationsStates

Let’sdiscussthechoiceswehave:1. Whatarethesetofdecisionsthepredictorneedstomake?2. Howdothesedecisionsinteract?Factors?3. Features?4. Learning?Inference?

Example1:Detectingobjectsandparts

7[Farhadi,etal]

Let’sdiscussthechoiceswehave:1. Whatarethesetofdecisionsthepredictorneedstomake?2. Howdothesedecisionsinteract?Factors?3. Features?4. Learning?Inference?

Example2:Informationextraction

PhilaeisaroboticEuropeanSpaceAgencylanderthataccompaniedtheRosettaspacecraftuntilitsdesignatedlandingonComet67P/Churyumov–Gerasimenko (67P),morethantenyearsafterdepartingEarth.On12November2014,thelanderachievedthefirst-evercontrolledtouchdownonacometnucleus.Itsinstrumentsareexpectedtoobtainthefirstimagesfromacomet'ssurfaceandmakethefirstinsituanalysistodetermineitscomposition.PhilaeistrackedandoperatedfromtheEuropeanSpaceOperationsCentre(ESOC)atDarmstadt,Germany.

8


Philae isaroboticEuropeanSpaceAgencylanderthataccompaniedtheRosettaspacecraftuntilits designatedlandingonComet67P/Churyumov–Gerasimenko (67P),morethantenyearsafterdepartingEarth.On12November2014,thelander achievedthefirst-evercontrolledtouchdownonacometnucleus.Its instrumentsareexpectedtoobtainthefirstimagesfromacomet'ssurfaceandmakethefirstinsituanalysistodetermineitscomposition.Philae istrackedandoperatedfromtheEuropeanSpaceOperationsCentre(ESOC)atDarmstadt,Germany.

9

Howdowemodelthisproblem?

TouchdownLander PhilaeDestination Comet 67PWhen? 12November2014



10

Philae



11

Philae Comet 67P



12

Philae Comet 67P

Touchdown

12 November 2014

Lander Dest. When



13

Howdowemodelthisproblem?

Philae Comet 67P

Touchdown

12 November 2014

Lander Dest. When



14

Philae Comet 67P

Touchdown

12 November 2014

Lander Dest. When Let’sdiscussthechoiceswehave:1. Whatarethesetofdecisionsthepredictor

needstomake?2. Howdothesedecisionsinteract?Factors?3. Features?4. Learning?Inference?

Modelingyourproblem

• Understandtheproblem:Whatshouldyourprogramproduce?– Istheredata?Veryoften,theanswerisno.L

• Whatarethedecisions/randomvariablesthatconstitutetheoutput?

• Howdotheyinteract?Identifyingfactors/parts– Someinteractionsarenatural,somearespurious(specifictoyoursmallcollectionofdata)– Someinteractionsmakeinferenceimpossibleforcomputationalreasons– Whatarethefeaturerepresentations?

• Learning– Shouldeveryscoringfunctionbejointlylearned?– Perhaps,learnsub-sectionsindependentlyandputthemtogetherwithinferenceattheend– Whichlearningalgorithm?

• Inference– Whatalgorithm?Howexpensiveisit?– Exactorapproximate?

15

Practical concerns in predicting structures...Practical concerns in predicting ... – We want to convert some input to an output – Model the conditional distribution of the output

Documents