learning important features through propagating activation ... · avanti shrikumar, peyton...

Learning Important Features Through PropagatingActivation Differences

Avanti Shrikumar Peyton Greenside Anshul Kundaje

Stanford University

ICML, 2017Presenter: Ritambhara Singh

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 1

/ 39

Outline

1 IntroductionMotivationBackgroundState-of-the-artDrawbacks

2 Proposed ApproachDeepLIFT MethodDefining ReferenceSolutionMultipliers and Chain RuleSeparating positive and negative contributionRules for assigning contributions

3 ResultsMNIST digit classificationDNA sequence classification

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 2

/ 39

Outline

1 IntroductionMotivationBackgroundState-of-the-artDrawbacks

2 Proposed ApproachDeepLIFT MethodDefining ReferenceSolutionMultipliers and Chain RuleSeparating positive and negative contributionRules for assigning contributions

3 ResultsMNIST digit classificationDNA sequence classification

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 3

/ 39

Motivation

Interpretabilty of neural networks :Assign importance score toinputs for a given output.

Importance is defined in terms of differences from a ‘reference’ state.

Propagates importance signal even when gradient is zero.

Gives separate consideration to positive and negative contributions.

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 4

/ 39

Motivation

Interpretabilty of neural networks :Assign importance score toinputs for a given output.

Importance is defined in terms of differences from a ‘reference’ state.

Propagates importance signal even when gradient is zero.

Gives separate consideration to positive and negative contributions.

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 4

/ 39

Motivation

Interpretabilty of neural networks :Assign importance score toinputs for a given output.

Importance is defined in terms of differences from a ‘reference’ state.

Propagates importance signal even when gradient is zero.

Gives separate consideration to positive and negative contributions.

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 4

/ 39

Motivation

Interpretabilty of neural networks :Assign importance score toinputs for a given output.

Importance is defined in terms of differences from a ‘reference’ state.

Propagates importance signal even when gradient is zero.

Gives separate consideration to positive and negative contributions.

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 4

/ 39

Outline

1 IntroductionMotivationBackgroundState-of-the-artDrawbacks

2 Proposed ApproachDeepLIFT MethodDefining ReferenceSolutionMultipliers and Chain RuleSeparating positive and negative contributionRules for assigning contributions

3 ResultsMNIST digit classificationDNA sequence classification

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 5

/ 39

Interpretation of Neural Networks

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 6

/ 39

Outline

1 IntroductionMotivationBackgroundState-of-the-artDrawbacks

2 Proposed ApproachDeepLIFT MethodDefining ReferenceSolutionMultipliers and Chain RuleSeparating positive and negative contributionRules for assigning contributions

3 ResultsMNIST digit classificationDNA sequence classification

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 7

/ 39

State-of-the-art

Perturbation-based forward propagation approaches: Zeiler andFergus (2013), Zhou and Troyanskaya (2015).

Backpropagation-based approaches: Saliency maps: Simonyan etal. (2013), Guided Backpropagation: Springenberg et al. (2014)

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 8

/ 39

Outline

1 IntroductionMotivationBackgroundState-of-the-artDrawbacks

2 Proposed ApproachDeepLIFT MethodDefining ReferenceSolutionMultipliers and Chain RuleSeparating positive and negative contributionRules for assigning contributions

3 ResultsMNIST digit classificationDNA sequence classification

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 9

/ 39

Saturation problem

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 10

/ 39

Saturation problem

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 11

/ 39

Thresholding Problem

y = max(0, x − 10)

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 12

/ 39

Outline

1 IntroductionMotivationBackgroundState-of-the-artDrawbacks

2 Proposed ApproachDeepLIFT MethodDefining ReferenceSolutionMultipliers and Chain RuleSeparating positive and negative contributionRules for assigning contributions

3 ResultsMNIST digit classificationDNA sequence classification

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 13

/ 39

Philosophy

Explains difference in output from some ‘reference’ output in terms ofdifference on input from some ‘reference’ input.

Summation-to-delta property:

Σni=1C∆xi∆t = ∆t (1)

Blame ∆t on ∆x1,∆x2, . . .

C∆xi∆t can be non-zero even when δtδxi

is zero.

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 14

/ 39

Philosophy

Explains difference in output from some ‘reference’ output in terms ofdifference on input from some ‘reference’ input.

Summation-to-delta property:

Σni=1C∆xi∆t = ∆t (1)

Blame ∆t on ∆x1,∆x2, . . .

C∆xi∆t can be non-zero even when δtδxi

is zero.

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 14

/ 39

Philosophy

Explains difference in output from some ‘reference’ output in terms ofdifference on input from some ‘reference’ input.

Summation-to-delta property:

Σni=1C∆xi∆t = ∆t (1)

Blame ∆t on ∆x1,∆x2, . . .

C∆xi∆t can be non-zero even when δtδxi

is zero.

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 14

/ 39

Philosophy

Explains difference in output from some ‘reference’ output in terms ofdifference on input from some ‘reference’ input.

Summation-to-delta property:

Σni=1C∆xi∆t = ∆t (1)

Blame ∆t on ∆x1,∆x2, . . .

C∆xi∆t can be non-zero even when δtδxi

is zero.

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 14

/ 39

Outline

1 IntroductionMotivationBackgroundState-of-the-artDrawbacks

2 Proposed ApproachDeepLIFT MethodDefining ReferenceSolutionMultipliers and Chain RuleSeparating positive and negative contributionRules for assigning contributions

3 ResultsMNIST digit classificationDNA sequence classification

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 15

/ 39

Defining Reference

Given neuron x with inputs i1, i2, . . . such that x = f (i1, i2, . . . )

Given reference activations i01 , i02 , . . . of the input:

x0 = f (i01 , i02 , . . . ) (2)

Choose reference input and propagate activations though the net.

Good reference will rely on domain knowledge: “What am I interestedin measuring difference against?”

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 16

/ 39

Outline

1 IntroductionMotivationBackgroundState-of-the-artDrawbacks

2 Proposed ApproachDeepLIFT MethodDefining ReferenceSolutionMultipliers and Chain RuleSeparating positive and negative contributionRules for assigning contributions

3 ResultsMNIST digit classificationDNA sequence classification

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 17

/ 39

Saturation Problem

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 18

/ 39

Thresholding Problem

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 19

/ 39

Outline

1 IntroductionMotivationBackgroundState-of-the-artDrawbacks

2 Proposed ApproachDeepLIFT MethodDefining ReferenceSolutionMultipliers and Chain RuleSeparating positive and negative contributionRules for assigning contributions

3 ResultsMNIST digit classificationDNA sequence classification

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 20

/ 39

Multipliers

m∆x∆t =C∆x∆t

∆t(3)

Multiplier is the contribution of ∆x to ∆t divided by ∆x

Compare: partial derivative = δtδx

Infinitesimal contribution of δx to δt, divided by δx

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 21

/ 39

Chain Rule

m∆xi∆z = Σjm∆xi∆yjm∆yj∆z (4)

Can be computed efficiently via backpropagation

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 22

/ 39

Outline

1 IntroductionMotivationBackgroundState-of-the-artDrawbacks

2 Proposed ApproachDeepLIFT MethodDefining ReferenceSolutionMultipliers and Chain RuleSeparating positive and negative contributionRules for assigning contributions

3 ResultsMNIST digit classificationDNA sequence classification

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 23

/ 39

Separating positive and negative contribution

In some cases, important to treat positive and negative contributionsdifferently.

Introduce ∆x+i and ∆x−i , such that:

∆xi = ∆x+i + ∆x−i ;C∆xi∆t = C∆x+

i ∆t + C∆x−i ∆t

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 24

/ 39

Outline

1 IntroductionMotivationBackgroundState-of-the-artDrawbacks

2 Proposed ApproachDeepLIFT MethodDefining ReferenceSolutionMultipliers and Chain RuleSeparating positive and negative contributionRules for assigning contributions

3 ResultsMNIST digit classificationDNA sequence classification

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 25

/ 39

Linear Rule

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 26

/ 39

Rescale Rule

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 27

/ 39

Where it works

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 28

/ 39

Where it works

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 29

/ 39

Where it fails: “min” (AND) relation

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 30

/ 39

Where it fails: “min” (AND) relation

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 31

/ 39

RevealCancel Rule

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 32

/ 39

Solution: “min” (AND) relation

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 33

/ 39

Outline

1 IntroductionMotivationBackgroundState-of-the-artDrawbacks

2 Proposed ApproachDeepLIFT MethodDefining ReferenceSolutionMultipliers and Chain RuleSeparating positive and negative contributionRules for assigning contributions

3 ResultsMNIST digit classificationDNA sequence classification

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 34

/ 39

MNIST digit classification

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 35

/ 39

Outline

1 IntroductionMotivationBackgroundState-of-the-artDrawbacks

2 Proposed ApproachDeepLIFT MethodDefining ReferenceSolutionMultipliers and Chain RuleSeparating positive and negative contributionRules for assigning contributions

3 ResultsMNIST digit classificationDNA sequence classification

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 36

/ 39

DNA sequence classification

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 37

/ 39

DNA sequence classification

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 38

/ 39

Summary

Novel approach for computing importance scores based on differencesfrom the ‘reference’.

Using difference-from-reference allows information to propagate evenwhen the gradient is zero

Separates contributions from positive and negative terms

Video at : https://www.youtube.com/watch?v=v8cxYjNZAXc&

index=1&list=PLJLjQOkqSRTP3cLB2cOOi_bQFw6KPGKML

Slides at: https://drive.google.com/file/d/0B15F_

QN41VQXbkVkcTVQYTVQNVE/view

Future Direction

Applying DeepLIFT to RNNsCompute ‘reference’ empirically from data

Avanti Shrikumar, Peyton Greenside, Anshul Kundaje (Stanford University)Learning Important Features Through Propagating Activation DifferencesICML, 2017 Presenter: Ritambhara Singh 39

/ 39