Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12....

46
CS839: Probabilistic Graphical Models Lecture 16: Spectral Algorithms for GMs Theo Rekatsinas 1

Transcript of Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12....

Page 1: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

CS839:ProbabilisticGraphicalModels

Lecture16:SpectralAlgorithmsforGMsTheoRekatsinas

1

Page 2: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

Overview

2

• AnoverviewoftheDLcomponents• Historicalremarks:earlydaysofneuralnetworks• Modernbuildingblocks:units,layers,activationsfunctions,lossfunctions,etc.• Reverse-modeautomaticdifferentiation(akabackpropagation)Distributedrepresentations

• SimilaritiesanddifferencesbetweenGMsandNNs• Graphicalmodelsvs.computationalgraphs• SigmoidBeliefNetworksasgraphicalmodels• DeepBeliefNetworksandBoltzmannMachines

• CombiningDLmethodsandGMs• UsingoutputsofNNsasinputstoGMs• GMswithpotentialfunctionsrepresentedbyNNs• NNswithstructuredoutputs

Page 3: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

History- Motivation

3

Page 4: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

PerceptronandNeuralNetworks

4

Page 5: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

ThePerceptronLearningAlgorithm

5

Page 6: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

ThePerceptronLearningAlgorithm

6

Page 7: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

NeuralNetworkModel

7

Page 8: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

Combinedlogisticmodels

8

Page 9: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

Combinedlogisticmodels

9

Page 10: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

Combinedlogisticmodels

10

Page 11: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

Notreally,notargetforhiddenunits...

11

Page 12: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

Backpropagation:Reverse-modedifferentiation

12

Page 13: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

Backpropagation:Reverse-modedifferentiation

13

Page 14: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

Modelbuildingblocks

14

Page 15: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

Modelbuildingblocks

15

Page 16: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

Buildingblocksofdeepnetworks

16

Page 17: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

Buildingblocksofdeepnetworks

17

Page 18: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

Hand-craftedfeatures

18

Page 19: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

Hand-craftedfeatures

19

Page 20: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

UsingDNNsforhierarchicalrepresentations

20

Page 21: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

21

Page 22: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

GraphicalmodelsvsDeepnets

22

Page 23: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

GraphicalmodelsvsDeepnets

23

Page 24: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

GraphicalmodelsvsDeepnets

24

Page 25: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

GraphicalmodelsvsDeepnets

25

Page 26: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

GraphicalmodelsvsDeepnets

26

Page 27: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

RestrictedBoltzmannMachines

27

Page 28: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

RestrictedBoltzmannMachines:LearningandInference

28

Page 29: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

RestrictedBoltzmannMachines:LearningandInference

29

Page 30: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

RestrictedBoltzmannMachines:LearningandInference

30

Page 31: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

SigmoidBeliefNetworks

31

Page 32: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

RBMsareinfinitebeliefnetworks

32

Page 33: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

RBMsareinfinitebeliefnetworks

33

Page 34: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

RBMsareinfinitebeliefnetworks

34

Page 35: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

RBMsareinfinitebeliefnetworks

35

Page 36: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

Deepbeliefnetworks:layer-wisepre-training

36

Page 37: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

DeepBoltzmannMachines

37

Page 38: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

DeepBoltzmannMachines

38

Page 39: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

39

GraphicalmodelsvsDeepnets

Page 40: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

40

CombiningsequentialNNsandGMs[Gravesetal.2013]

Page 41: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

41

CombiningsequentialNNsandGMs[Gravesetal.2013]

Page 42: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

42

HybridNNsandconditionalGMs

Page 43: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

43

HybridNNsandconditionalGMs

Page 44: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

44

HybridNNsandconditionalGMs

Page 45: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

45

Dealingwithstructuredprediction[Domke 2012]

Page 46: Lecture 16: Spectral Algorithms for GMs · Backpropagation: Reverse-mode differentiation 12. Backpropagation: Reverse-mode differentiation 13. Model building blocks 14. Model building

Summary

46

• DL&GM:thefieldsaresimilarinthebeginning(structure,energy,etc.),andthendivergetotheirownsignaturepipelines• DL:mosteffortisdirectedtocomparingdifferentarchitecturesandtheircomponents(basedonempiricalperformanceonadownstreamtask)• DLmodelsaregoodatlearningrobusthierarchicalrepresentationsfromthedataandsuitableforsimplereasoning(“low-levelcognition”)

• GM:lotsofeffortsaredirectedtoimprovinginferenceaccuracyandconvergencespeed• GMsarebestforprovablycorrectinferenceandsuitableforhigh-levelcomplexreasoningtasks(“high-levelcognition”)

• Convergenceofbothfieldsisverypromising!