SlideShare una empresa de Scribd logo
1 de 17
Terry Taewoong Um (terry.t.um@gmail.com)
University of Waterloo
Department of Electrical & Computer Engineering
Terry T. Um
UNDERSTANDING BLACK-BOX PRED
-ICTION VIA INFLUENCE FUNCTIONS
1
TODAY’S PAPER
Terry Taewoong Um (terry.t.um@gmail.com)
ICML2017 best paper
https://youtu.be/0w9fLX_T6tY
QUESTIONS
Terry Taewoong Um (terry.t.um@gmail.com)
• How can we explain the predictions of a black-box model?
• Why did the system make this prediction?
• How can we explain where the model came from?
• What would happen if the values of a training point where
slightly changed?
INTERPRETATION OF DL RESULTS
Terry Taewoong Um (terry.t.um@gmail.com)
• Retrieving images that maximally activate a neuron [Girshick et al. 2014]
• Finding the most influential part from the image [Zhou et al. 2016]
• Learning a simpler model around a test point [Ribeiro et al. 2016]
But, they assumed a
fixed model
 My NN is a function
of training inputs
INFLUENCE OF A TRAINING POINT
Terry Taewoong Um (terry.t.um@gmail.com)
• What is the influence of a training example for
the model (or for the loss of a test example)?
Optimal model param. :
Model param. by training w/o z :
Model param. by upweighting z :
without z == (𝜖 = −
1
𝑛
)
• The influence of upweighting z on the parameters 𝜃
INFLUENCE OF A TRAINING POINT
• Influence vs. Euclidean distance
INFLUENCE OF A TRAINING POINT
Terry Taewoong Um (terry.t.um@gmail.com)
• The influence of upweighting z on the parameters 𝜃
• The influence of upweighting z on the loss at a test point
PERTURBING A TRAINING POINT
Terry Taewoong Um (terry.t.um@gmail.com)
• Move 𝜖 mass from 𝑧 to 𝑧 𝛿
• If x is continuous and 𝛿 is small
• The effect of 𝑧  𝑧 𝛿 on the loss at a test point
SUMMARY
Terry Taewoong Um (terry.t.um@gmail.com)
• The influence of 𝑧  𝑧 𝛿 on the loss at a test point
• The influence of upweighting z on the parameters 𝜃
• The influence of upweighting z on the loss at a test point
EXAMPLE
Terry Taewoong Um (terry.t.um@gmail.com)
• The influence of upweighting z
• In logistic regression,
• Test : 7, Train : 7 (green), 1 (red)
SEVERAL PROBLEMS
Terry Taewoong Um (terry.t.um@gmail.com)
• Calculation of
 Use Hessian-vector products (HVPs)

precompute 𝑠𝑡𝑒𝑠𝑡 by optimizing
or sampling-based approximation
SEVERAL PROBLEMS
Terry Taewoong Um (terry.t.um@gmail.com)
• What if is non-convex, so H < 0
 Assuming that is a local minimum point, define a quadratic loss
Then calculate using the above
 empirically working!
• Influence function vs. retraining
SEVERAL PROBLEMS
Terry Taewoong Um (terry.t.um@gmail.com)
• What if is non-differentiable?
e.g.) hinge loss
 Use a differentiable variation of the hinge loss
APPLICATIONS
Terry Taewoong Um (terry.t.um@gmail.com)
• Understanding model behavior
APPLICATIONS
Terry Taewoong Um (terry.t.um@gmail.com)
• Adversarial examples
c.f.) The effect of 𝑧  𝑧 𝛿 on the loss at a test point
APPLICATIONS
Terry Taewoong Um (terry.t.um@gmail.com)
• Debugging domain mismatch
APPLICATIONS
Terry Taewoong Um (terry.t.um@gmail.com)
• Fixing mislabeled examples

Más contenido relacionado

La actualidad más candente

La actualidad más candente (8)

Deep Variational Bayes Filters (2017)
Deep Variational Bayes Filters (2017)Deep Variational Bayes Filters (2017)
Deep Variational Bayes Filters (2017)
 
Introduction to Deep Learning with TensorFlow
Introduction to Deep Learning with TensorFlowIntroduction to Deep Learning with TensorFlow
Introduction to Deep Learning with TensorFlow
 
On Calibration of Modern Neural Networks (2017)
On Calibration of Modern Neural Networks (2017)On Calibration of Modern Neural Networks (2017)
On Calibration of Modern Neural Networks (2017)
 
Tabu search
Tabu searchTabu search
Tabu search
 
Novel Machine Learning Methods for Extraction of Features Characterizing Data...
Novel Machine Learning Methods for Extraction of Features Characterizing Data...Novel Machine Learning Methods for Extraction of Features Characterizing Data...
Novel Machine Learning Methods for Extraction of Features Characterizing Data...
 
Higher Order Fused Regularization for Supervised Learning with Grouped Parame...
Higher Order Fused Regularization for Supervised Learning with Grouped Parame...Higher Order Fused Regularization for Supervised Learning with Grouped Parame...
Higher Order Fused Regularization for Supervised Learning with Grouped Parame...
 
Joint contrastive learning with infinite possibilities
Joint contrastive learning with infinite possibilitiesJoint contrastive learning with infinite possibilities
Joint contrastive learning with infinite possibilities
 
A Reinforcement Learning Approach for Hybrid Flexible Flowline Scheduling Pro...
A Reinforcement Learning Approach for Hybrid Flexible Flowline Scheduling Pro...A Reinforcement Learning Approach for Hybrid Flexible Flowline Scheduling Pro...
A Reinforcement Learning Approach for Hybrid Flexible Flowline Scheduling Pro...
 

Destacado

Destacado (20)

Lie Group Formulation for Robot Mechanics
Lie Group Formulation for Robot MechanicsLie Group Formulation for Robot Mechanics
Lie Group Formulation for Robot Mechanics
 
기계학습 / 딥러닝이란 무엇인가
기계학습 / 딥러닝이란 무엇인가기계학습 / 딥러닝이란 무엇인가
기계학습 / 딥러닝이란 무엇인가
 
[모두의연구소] 쫄지말자딥러닝
[모두의연구소] 쫄지말자딥러닝[모두의연구소] 쫄지말자딥러닝
[모두의연구소] 쫄지말자딥러닝
 
인공 신경망 구현에 관한 간단한 설명
인공 신경망 구현에 관한 간단한 설명인공 신경망 구현에 관한 간단한 설명
인공 신경망 구현에 관한 간단한 설명
 
R 프로그래밍 기본 문법
R 프로그래밍 기본 문법R 프로그래밍 기본 문법
R 프로그래밍 기본 문법
 
머신 러닝 입문 #1-머신러닝 소개와 kNN 소개
머신 러닝 입문 #1-머신러닝 소개와 kNN 소개머신 러닝 입문 #1-머신러닝 소개와 kNN 소개
머신 러닝 입문 #1-머신러닝 소개와 kNN 소개
 
쫄지말자딥러닝2 - CNN RNN 포함버전
쫄지말자딥러닝2 - CNN RNN 포함버전쫄지말자딥러닝2 - CNN RNN 포함버전
쫄지말자딥러닝2 - CNN RNN 포함버전
 
인공지능, 기계학습 그리고 딥러닝
인공지능, 기계학습 그리고 딥러닝인공지능, 기계학습 그리고 딥러닝
인공지능, 기계학습 그리고 딥러닝
 
Large Scale Deep Learning with TensorFlow
Large Scale Deep Learning with TensorFlow Large Scale Deep Learning with TensorFlow
Large Scale Deep Learning with TensorFlow
 
U2 product For Wiseeco
U2 product For WiseecoU2 product For Wiseeco
U2 product For Wiseeco
 
GStreamer-VAAPI: Hardware-accelerated encoding and decoding on Intel hardware...
GStreamer-VAAPI: Hardware-accelerated encoding and decoding on Intel hardware...GStreamer-VAAPI: Hardware-accelerated encoding and decoding on Intel hardware...
GStreamer-VAAPI: Hardware-accelerated encoding and decoding on Intel hardware...
 
Global mobile market report
Global mobile market reportGlobal mobile market report
Global mobile market report
 
알파고 (바둑 인공지능)의 작동 원리
알파고 (바둑 인공지능)의 작동 원리알파고 (바둑 인공지능)의 작동 원리
알파고 (바둑 인공지능)의 작동 원리
 
Docker 로 Linux 없이 Linux 환경에서 개발하기
Docker 로 Linux 없이 Linux 환경에서 개발하기Docker 로 Linux 없이 Linux 환경에서 개발하기
Docker 로 Linux 없이 Linux 환경에서 개발하기
 
Pure Function and Honest Design
Pure Function and Honest DesignPure Function and Honest Design
Pure Function and Honest Design
 
Pitfalls of Object Oriented Programming by SONY
Pitfalls of Object Oriented Programming by SONYPitfalls of Object Oriented Programming by SONY
Pitfalls of Object Oriented Programming by SONY
 
2017 k8s and OpenStack-Helm
2017 k8s and OpenStack-Helm2017 k8s and OpenStack-Helm
2017 k8s and OpenStack-Helm
 
1, 빅데이터 시대의 인공지능 문동선 v2
1, 빅데이터 시대의 인공지능 문동선 v21, 빅데이터 시대의 인공지능 문동선 v2
1, 빅데이터 시대의 인공지능 문동선 v2
 
클라우드 네트워킹과 SDN 그리고 OpenStack
클라우드 네트워킹과 SDN 그리고 OpenStack클라우드 네트워킹과 SDN 그리고 OpenStack
클라우드 네트워킹과 SDN 그리고 OpenStack
 
java 8 람다식 소개와 의미 고찰
java 8 람다식 소개와 의미 고찰java 8 람다식 소개와 의미 고찰
java 8 람다식 소개와 의미 고찰
 

Más de Terry Taewoong Um

Más de Terry Taewoong Um (6)

#44. KAIST에서 "대학 유죄"를 외치다: ART Lab의 도전
#44. KAIST에서 "대학 유죄"를 외치다: ART Lab의 도전#44. KAIST에서 "대학 유죄"를 외치다: ART Lab의 도전
#44. KAIST에서 "대학 유죄"를 외치다: ART Lab의 도전
 
A brief introduction to OCR (Optical character recognition)
A brief introduction to OCR (Optical character recognition)A brief introduction to OCR (Optical character recognition)
A brief introduction to OCR (Optical character recognition)
 
인공지능의 사회정의의 편이 될 수 있을까? (인공지능과 법)
인공지능의 사회정의의 편이 될 수 있을까? (인공지능과 법)인공지능의 사회정의의 편이 될 수 있을까? (인공지능과 법)
인공지능의 사회정의의 편이 될 수 있을까? (인공지능과 법)
 
Deep learning (Machine learning) tutorial for beginners
Deep learning (Machine learning) tutorial for beginnersDeep learning (Machine learning) tutorial for beginners
Deep learning (Machine learning) tutorial for beginners
 
Deep Learning: A Critical Appraisal (2018)
Deep Learning: A Critical Appraisal (2018)Deep Learning: A Critical Appraisal (2018)
Deep Learning: A Critical Appraisal (2018)
 
로봇과 인공지능, 그리고 미래의 노동
로봇과 인공지능, 그리고 미래의 노동로봇과 인공지능, 그리고 미래의 노동
로봇과 인공지능, 그리고 미래의 노동
 

Último

DeepFakes presentation : brief idea of DeepFakes
DeepFakes presentation : brief idea of DeepFakesDeepFakes presentation : brief idea of DeepFakes
DeepFakes presentation : brief idea of DeepFakes
MayuraD1
 
Cara Menggugurkan Sperma Yang Masuk Rahim Biyar Tidak Hamil
Cara Menggugurkan Sperma Yang Masuk Rahim Biyar Tidak HamilCara Menggugurkan Sperma Yang Masuk Rahim Biyar Tidak Hamil
Cara Menggugurkan Sperma Yang Masuk Rahim Biyar Tidak Hamil
Cara Menggugurkan Kandungan 087776558899
 
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
ssuser89054b
 

Último (20)

DeepFakes presentation : brief idea of DeepFakes
DeepFakes presentation : brief idea of DeepFakesDeepFakes presentation : brief idea of DeepFakes
DeepFakes presentation : brief idea of DeepFakes
 
DC MACHINE-Motoring and generation, Armature circuit equation
DC MACHINE-Motoring and generation, Armature circuit equationDC MACHINE-Motoring and generation, Armature circuit equation
DC MACHINE-Motoring and generation, Armature circuit equation
 
Thermal Engineering-R & A / C - unit - V
Thermal Engineering-R & A / C - unit - VThermal Engineering-R & A / C - unit - V
Thermal Engineering-R & A / C - unit - V
 
Thermal Engineering Unit - I & II . ppt
Thermal Engineering  Unit - I & II . pptThermal Engineering  Unit - I & II . ppt
Thermal Engineering Unit - I & II . ppt
 
Navigating Complexity: The Role of Trusted Partners and VIAS3D in Dassault Sy...
Navigating Complexity: The Role of Trusted Partners and VIAS3D in Dassault Sy...Navigating Complexity: The Role of Trusted Partners and VIAS3D in Dassault Sy...
Navigating Complexity: The Role of Trusted Partners and VIAS3D in Dassault Sy...
 
Moment Distribution Method For Btech Civil
Moment Distribution Method For Btech CivilMoment Distribution Method For Btech Civil
Moment Distribution Method For Btech Civil
 
GEAR TRAIN- BASIC CONCEPTS AND WORKING PRINCIPLE
GEAR TRAIN- BASIC CONCEPTS AND WORKING PRINCIPLEGEAR TRAIN- BASIC CONCEPTS AND WORKING PRINCIPLE
GEAR TRAIN- BASIC CONCEPTS AND WORKING PRINCIPLE
 
Bhubaneswar🌹Call Girls Bhubaneswar ❤Komal 9777949614 💟 Full Trusted CALL GIRL...
Bhubaneswar🌹Call Girls Bhubaneswar ❤Komal 9777949614 💟 Full Trusted CALL GIRL...Bhubaneswar🌹Call Girls Bhubaneswar ❤Komal 9777949614 💟 Full Trusted CALL GIRL...
Bhubaneswar🌹Call Girls Bhubaneswar ❤Komal 9777949614 💟 Full Trusted CALL GIRL...
 
Unit 4_Part 1 CSE2001 Exception Handling and Function Template and Class Temp...
Unit 4_Part 1 CSE2001 Exception Handling and Function Template and Class Temp...Unit 4_Part 1 CSE2001 Exception Handling and Function Template and Class Temp...
Unit 4_Part 1 CSE2001 Exception Handling and Function Template and Class Temp...
 
Cara Menggugurkan Sperma Yang Masuk Rahim Biyar Tidak Hamil
Cara Menggugurkan Sperma Yang Masuk Rahim Biyar Tidak HamilCara Menggugurkan Sperma Yang Masuk Rahim Biyar Tidak Hamil
Cara Menggugurkan Sperma Yang Masuk Rahim Biyar Tidak Hamil
 
kiln thermal load.pptx kiln tgermal load
kiln thermal load.pptx kiln tgermal loadkiln thermal load.pptx kiln tgermal load
kiln thermal load.pptx kiln tgermal load
 
A CASE STUDY ON CERAMIC INDUSTRY OF BANGLADESH.pptx
A CASE STUDY ON CERAMIC INDUSTRY OF BANGLADESH.pptxA CASE STUDY ON CERAMIC INDUSTRY OF BANGLADESH.pptx
A CASE STUDY ON CERAMIC INDUSTRY OF BANGLADESH.pptx
 
S1S2 B.Arch MGU - HOA1&2 Module 3 -Temple Architecture of Kerala.pptx
S1S2 B.Arch MGU - HOA1&2 Module 3 -Temple Architecture of Kerala.pptxS1S2 B.Arch MGU - HOA1&2 Module 3 -Temple Architecture of Kerala.pptx
S1S2 B.Arch MGU - HOA1&2 Module 3 -Temple Architecture of Kerala.pptx
 
Computer Networks Basics of Network Devices
Computer Networks  Basics of Network DevicesComputer Networks  Basics of Network Devices
Computer Networks Basics of Network Devices
 
Work-Permit-Receiver-in-Saudi-Aramco.pptx
Work-Permit-Receiver-in-Saudi-Aramco.pptxWork-Permit-Receiver-in-Saudi-Aramco.pptx
Work-Permit-Receiver-in-Saudi-Aramco.pptx
 
Online electricity billing project report..pdf
Online electricity billing project report..pdfOnline electricity billing project report..pdf
Online electricity billing project report..pdf
 
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
 
Thermal Engineering -unit - III & IV.ppt
Thermal Engineering -unit - III & IV.pptThermal Engineering -unit - III & IV.ppt
Thermal Engineering -unit - III & IV.ppt
 
Employee leave management system project.
Employee leave management system project.Employee leave management system project.
Employee leave management system project.
 
HAND TOOLS USED AT ELECTRONICS WORK PRESENTED BY KOUSTAV SARKAR
HAND TOOLS USED AT ELECTRONICS WORK PRESENTED BY KOUSTAV SARKARHAND TOOLS USED AT ELECTRONICS WORK PRESENTED BY KOUSTAV SARKAR
HAND TOOLS USED AT ELECTRONICS WORK PRESENTED BY KOUSTAV SARKAR
 

Understanding Black-box Predictions via Influence Functions (2017)

  • 1. Terry Taewoong Um (terry.t.um@gmail.com) University of Waterloo Department of Electrical & Computer Engineering Terry T. Um UNDERSTANDING BLACK-BOX PRED -ICTION VIA INFLUENCE FUNCTIONS 1
  • 2. TODAY’S PAPER Terry Taewoong Um (terry.t.um@gmail.com) ICML2017 best paper https://youtu.be/0w9fLX_T6tY
  • 3. QUESTIONS Terry Taewoong Um (terry.t.um@gmail.com) • How can we explain the predictions of a black-box model? • Why did the system make this prediction? • How can we explain where the model came from? • What would happen if the values of a training point where slightly changed?
  • 4. INTERPRETATION OF DL RESULTS Terry Taewoong Um (terry.t.um@gmail.com) • Retrieving images that maximally activate a neuron [Girshick et al. 2014] • Finding the most influential part from the image [Zhou et al. 2016] • Learning a simpler model around a test point [Ribeiro et al. 2016] But, they assumed a fixed model  My NN is a function of training inputs
  • 5. INFLUENCE OF A TRAINING POINT Terry Taewoong Um (terry.t.um@gmail.com) • What is the influence of a training example for the model (or for the loss of a test example)? Optimal model param. : Model param. by training w/o z : Model param. by upweighting z : without z == (𝜖 = − 1 𝑛 ) • The influence of upweighting z on the parameters 𝜃
  • 6. INFLUENCE OF A TRAINING POINT • Influence vs. Euclidean distance
  • 7. INFLUENCE OF A TRAINING POINT Terry Taewoong Um (terry.t.um@gmail.com) • The influence of upweighting z on the parameters 𝜃 • The influence of upweighting z on the loss at a test point
  • 8. PERTURBING A TRAINING POINT Terry Taewoong Um (terry.t.um@gmail.com) • Move 𝜖 mass from 𝑧 to 𝑧 𝛿 • If x is continuous and 𝛿 is small • The effect of 𝑧  𝑧 𝛿 on the loss at a test point
  • 9. SUMMARY Terry Taewoong Um (terry.t.um@gmail.com) • The influence of 𝑧  𝑧 𝛿 on the loss at a test point • The influence of upweighting z on the parameters 𝜃 • The influence of upweighting z on the loss at a test point
  • 10. EXAMPLE Terry Taewoong Um (terry.t.um@gmail.com) • The influence of upweighting z • In logistic regression, • Test : 7, Train : 7 (green), 1 (red)
  • 11. SEVERAL PROBLEMS Terry Taewoong Um (terry.t.um@gmail.com) • Calculation of  Use Hessian-vector products (HVPs)  precompute 𝑠𝑡𝑒𝑠𝑡 by optimizing or sampling-based approximation
  • 12. SEVERAL PROBLEMS Terry Taewoong Um (terry.t.um@gmail.com) • What if is non-convex, so H < 0  Assuming that is a local minimum point, define a quadratic loss Then calculate using the above  empirically working! • Influence function vs. retraining
  • 13. SEVERAL PROBLEMS Terry Taewoong Um (terry.t.um@gmail.com) • What if is non-differentiable? e.g.) hinge loss  Use a differentiable variation of the hinge loss
  • 14. APPLICATIONS Terry Taewoong Um (terry.t.um@gmail.com) • Understanding model behavior
  • 15. APPLICATIONS Terry Taewoong Um (terry.t.um@gmail.com) • Adversarial examples c.f.) The effect of 𝑧  𝑧 𝛿 on the loss at a test point
  • 16. APPLICATIONS Terry Taewoong Um (terry.t.um@gmail.com) • Debugging domain mismatch
  • 17. APPLICATIONS Terry Taewoong Um (terry.t.um@gmail.com) • Fixing mislabeled examples