SlideShare a Scribd company logo
10
Most read
15
Most read
18
Most read
Why should I
trust you ?
Explaining the prediction of
any classifiers
Presented by
2
https://www.facebook.com/cannyedge/
Reference Paper
○ Originally announced in
February 2016.
○ University of Washington
○ Cited by 2054
3
https://arxiv.org/pdf/1602.04938.pdf
4
Sometime we don’t know if we can
trust ML prediction …
Or
If you can trust a ML model
Alternatives is to use only interpretable models
❖ Naive Bayes Classifier
❖ K-Nearest Neighbors
❖ Linear regression
❖ Logistic regression
❖ Decision trees
5
Predictive performance is LOST
compared to other complex machine
learning models
6
What if we could understand
why ML making prediction ?
7
It’s easier to trust a
prediction if you understand
the reasons for it?
8
Or figure out when you
shouldn’t trust the model?
I’m sure it’s 100% WOLF
Model was
predicting snow not
wolf
We can’t trust the
model even if the
accuracy was high !!
With LIME algorithm you can!
10
11
Mathematical expression is
The recipe for training
local surrogate models:
1. Select your instance of interest
2. Perturb your dataset and get the black box predictions for these new
points.
3. Weight the new samples according to their proximity to the instance
of interest.
4. Train a weighted, interpretable model on the dataset with the
variations.
5. Explain the prediction by interpreting the local model.
12
13
14
Example: Deep networks for image
GoogLeNet
TOP 3 Classes
“Electric Guitar” (p = 0.32)
“Acoustic guitar” (p = 0.24)
“Labrador” (p = 0.21)
15
Example: Deep networks for image
Explaining an image classification by Google’s Inception neural network.
The top 3 classes predicted are
“Electric Guitar” (p = 0.32),
“Acoustic guitar” (p = 0.24)
“Labrador” (p = 0.21)
○ Sampling data rely on Gaussian
distribution but not aware of
correlation between features
○ Have to define complexity in
advance
○ Instability of explanation
○ It’s possible that a linear model
might not be powerful enough
to explain
Advantage vs Disadvantage
○ Any Complex models with
single Interpretable
○ Easy, Concise, Human-friendly
Explanation
○ Can be applied for Tabular,
Text, and Image
○ Easy to implement in R and
Python
○ Different & Less Features than
Original models
16
17
LIME helps user in:
● Choosing between models
● Detect and improve
untrustworthy models
● Get insights into a model
HELP TO ANSWER:
Why should I trust you
Thank
You
Ref:
https://christophm.github.io/interpretable-ml-book/
https://arxiv.org/abs/1602.04938
18

More Related Content

PDF
Rで学ぶ『構造型モデル de 倒産確率推定』
PPTX
エンジニアライクにドキュメントを作成してみた件
PPTX
トランザクションをSerializableにする4つの方法
PDF
Newman アルゴリズムによるソーシャルグラフのクラスタリング
PPTX
Raspberry Piで始める自宅サーバー超入門
PDF
Apache Kuduは何がそんなに「速い」DBなのか? #dbts2017
PPTX
[DL輪読会]SafePicking: Learning Safe Object Extraction via Object-Level Mapping ...
PDF
OpenTelemetryを用いたObservability基礎の実装 with AWS Distro for OpenTelemetry(Kuberne...
Rで学ぶ『構造型モデル de 倒産確率推定』
エンジニアライクにドキュメントを作成してみた件
トランザクションをSerializableにする4つの方法
Newman アルゴリズムによるソーシャルグラフのクラスタリング
Raspberry Piで始める自宅サーバー超入門
Apache Kuduは何がそんなに「速い」DBなのか? #dbts2017
[DL輪読会]SafePicking: Learning Safe Object Extraction via Object-Level Mapping ...
OpenTelemetryを用いたObservability基礎の実装 with AWS Distro for OpenTelemetry(Kuberne...

What's hot (20)

PDF
NoOps?よろしいならば戦争だ
PDF
静的型付け言語Python
PPTX
顔認識アルゴリズム:Constrained local model を調べてみた
PDF
モジュールの凝集度・結合度・インタフェース
PPTX
ニューラルチューリングマシン入門
PDF
Introduction to YOLO detection model
PDF
新版 OutOfMemoryErrorを知る
PDF
Apache Atlasの現状とデータガバナンス事例 #hadoopreading
PDF
動画像理解のための深層学習アプローチ Deep learning approaches to video understanding
PDF
SSII2022 [OS3-01] 深層学習のための効率的なデータ収集と活用
PDF
[Postgre sql9.4新機能]レプリケーション・スロットの活用
PDF
Jiraの紹介(redmineとの比較視点にて)
PDF
Data-Centric AIの紹介
PDF
教師なし画像特徴表現学習の動向 {Un, Self} supervised representation learning (CVPR 2018 完全読破...
PDF
Ml system in_python
PDF
Mean Teacher
PDF
ゼロから始める転移学習
PPTX
cvsaisentan20141004 kanezaki
PPTX
210508 legaltech
PPTX
Oracle Database Vaultのご紹介
NoOps?よろしいならば戦争だ
静的型付け言語Python
顔認識アルゴリズム:Constrained local model を調べてみた
モジュールの凝集度・結合度・インタフェース
ニューラルチューリングマシン入門
Introduction to YOLO detection model
新版 OutOfMemoryErrorを知る
Apache Atlasの現状とデータガバナンス事例 #hadoopreading
動画像理解のための深層学習アプローチ Deep learning approaches to video understanding
SSII2022 [OS3-01] 深層学習のための効率的なデータ収集と活用
[Postgre sql9.4新機能]レプリケーション・スロットの活用
Jiraの紹介(redmineとの比較視点にて)
Data-Centric AIの紹介
教師なし画像特徴表現学習の動向 {Un, Self} supervised representation learning (CVPR 2018 完全読破...
Ml system in_python
Mean Teacher
ゼロから始める転移学習
cvsaisentan20141004 kanezaki
210508 legaltech
Oracle Database Vaultのご紹介
Ad

Similar to Why should i trust you summary paper (20)

PPTX
Practical Tips for Interpreting Machine Learning Models - Patrick Hall, H2O.ai
PDF
PPTX
Interpretable Machine Learning
PDF
Explainable AI - making ML and DL models more interpretable
PDF
Learning to learn Model Behavior: How to use "human-in-the-loop" to explain d...
PDF
Unsupervised Deep Learning (D2L1 Insight@DCU Machine Learning Workshop 2017)
PPTX
Machine Learning basics
PPTX
Explainable Machine Learning (Explainable ML)
PPTX
Machine Learning Interpretability - Mateusz Dymczyk - H2O AI World London 2018
PDF
Explore, Explain, and Debug aka Interpretable Machine Learning
PDF
Learning to Learn Model Behavior ( Capital One: data intelligence conference )
PDF
ML.pdf
PPTX
Graphs for Ai and ML
PDF
WIA 2019 - Steering Model Selection with Visual Diagnostics
PDF
Steering Model Selection with Visual Diagnostics: Women in Analytics 2019
PDF
Model evaluation in the land of deep learning
PDF
Human-Centered Interpretable Machine Learning
PDF
Steering Model Selection with Visual Diagnostics
PDF
AWS Certified Machine Learning Specialty
PDF
Model Evaluation in the land of Deep Learning
Practical Tips for Interpreting Machine Learning Models - Patrick Hall, H2O.ai
Interpretable Machine Learning
Explainable AI - making ML and DL models more interpretable
Learning to learn Model Behavior: How to use "human-in-the-loop" to explain d...
Unsupervised Deep Learning (D2L1 Insight@DCU Machine Learning Workshop 2017)
Machine Learning basics
Explainable Machine Learning (Explainable ML)
Machine Learning Interpretability - Mateusz Dymczyk - H2O AI World London 2018
Explore, Explain, and Debug aka Interpretable Machine Learning
Learning to Learn Model Behavior ( Capital One: data intelligence conference )
ML.pdf
Graphs for Ai and ML
WIA 2019 - Steering Model Selection with Visual Diagnostics
Steering Model Selection with Visual Diagnostics: Women in Analytics 2019
Model evaluation in the land of deep learning
Human-Centered Interpretable Machine Learning
Steering Model Selection with Visual Diagnostics
AWS Certified Machine Learning Specialty
Model Evaluation in the land of Deep Learning
Ad

Recently uploaded (20)

PPT
ISS -ESG Data flows What is ESG and HowHow
PDF
Business Analytics and business intelligence.pdf
PDF
Optimise Shopper Experiences with a Strong Data Estate.pdf
PPTX
A Complete Guide to Streamlining Business Processes
PPTX
STERILIZATION AND DISINFECTION-1.ppthhhbx
PDF
Global Data and Analytics Market Outlook Report
PDF
Data Engineering Interview Questions & Answers Cloud Data Stacks (AWS, Azure,...
PPTX
New ISO 27001_2022 standard and the changes
PPTX
FMIS 108 and AISlaudon_mis17_ppt_ch11.pptx
PDF
Jean-Georges Perrin - Spark in Action, Second Edition (2020, Manning Publicat...
PDF
Votre score augmente si vous choisissez une catégorie et que vous rédigez une...
PPTX
Leprosy and NLEP programme community medicine
PPTX
QUANTUM_COMPUTING_AND_ITS_POTENTIAL_APPLICATIONS[2].pptx
PDF
Navigating the Thai Supplements Landscape.pdf
PDF
Data Engineering Interview Questions & Answers Batch Processing (Spark, Hadoo...
PPTX
(Ali Hamza) Roll No: (F24-BSCS-1103).pptx
PPTX
Introduction to Inferential Statistics.pptx
PPTX
AI Strategy room jwfjksfksfjsjsjsjsjfsjfsj
PDF
Microsoft 365 products and services descrption
PPTX
Pilar Kemerdekaan dan Identi Bangsa.pptx
ISS -ESG Data flows What is ESG and HowHow
Business Analytics and business intelligence.pdf
Optimise Shopper Experiences with a Strong Data Estate.pdf
A Complete Guide to Streamlining Business Processes
STERILIZATION AND DISINFECTION-1.ppthhhbx
Global Data and Analytics Market Outlook Report
Data Engineering Interview Questions & Answers Cloud Data Stacks (AWS, Azure,...
New ISO 27001_2022 standard and the changes
FMIS 108 and AISlaudon_mis17_ppt_ch11.pptx
Jean-Georges Perrin - Spark in Action, Second Edition (2020, Manning Publicat...
Votre score augmente si vous choisissez une catégorie et que vous rédigez une...
Leprosy and NLEP programme community medicine
QUANTUM_COMPUTING_AND_ITS_POTENTIAL_APPLICATIONS[2].pptx
Navigating the Thai Supplements Landscape.pdf
Data Engineering Interview Questions & Answers Batch Processing (Spark, Hadoo...
(Ali Hamza) Roll No: (F24-BSCS-1103).pptx
Introduction to Inferential Statistics.pptx
AI Strategy room jwfjksfksfjsjsjsjsjfsjfsj
Microsoft 365 products and services descrption
Pilar Kemerdekaan dan Identi Bangsa.pptx

Why should i trust you summary paper

  • 1. Why should I trust you ? Explaining the prediction of any classifiers
  • 3. Reference Paper ○ Originally announced in February 2016. ○ University of Washington ○ Cited by 2054 3 https://arxiv.org/pdf/1602.04938.pdf
  • 4. 4 Sometime we don’t know if we can trust ML prediction … Or If you can trust a ML model
  • 5. Alternatives is to use only interpretable models ❖ Naive Bayes Classifier ❖ K-Nearest Neighbors ❖ Linear regression ❖ Logistic regression ❖ Decision trees 5 Predictive performance is LOST compared to other complex machine learning models
  • 6. 6 What if we could understand why ML making prediction ?
  • 7. 7 It’s easier to trust a prediction if you understand the reasons for it?
  • 8. 8 Or figure out when you shouldn’t trust the model? I’m sure it’s 100% WOLF Model was predicting snow not wolf We can’t trust the model even if the accuracy was high !!
  • 10. 10
  • 12. The recipe for training local surrogate models: 1. Select your instance of interest 2. Perturb your dataset and get the black box predictions for these new points. 3. Weight the new samples according to their proximity to the instance of interest. 4. Train a weighted, interpretable model on the dataset with the variations. 5. Explain the prediction by interpreting the local model. 12
  • 13. 13
  • 14. 14 Example: Deep networks for image GoogLeNet TOP 3 Classes “Electric Guitar” (p = 0.32) “Acoustic guitar” (p = 0.24) “Labrador” (p = 0.21)
  • 15. 15 Example: Deep networks for image Explaining an image classification by Google’s Inception neural network. The top 3 classes predicted are “Electric Guitar” (p = 0.32), “Acoustic guitar” (p = 0.24) “Labrador” (p = 0.21)
  • 16. ○ Sampling data rely on Gaussian distribution but not aware of correlation between features ○ Have to define complexity in advance ○ Instability of explanation ○ It’s possible that a linear model might not be powerful enough to explain Advantage vs Disadvantage ○ Any Complex models with single Interpretable ○ Easy, Concise, Human-friendly Explanation ○ Can be applied for Tabular, Text, and Image ○ Easy to implement in R and Python ○ Different & Less Features than Original models 16
  • 17. 17 LIME helps user in: ● Choosing between models ● Detect and improve untrustworthy models ● Get insights into a model HELP TO ANSWER: Why should I trust you

Editor's Notes

  • #12: ให้ g เป็นโมเดล(Linear Regression)ที่พยายามอธิบาย instance x ที่ต้องการ minimize loss function (MSE) ของ Original Model f (XGboost) โดยที่พยายาม keep ความซับซ้อนให้ต่ำทีสุด Omega (จำนวนfeatures ) และ G คือ กลุ่มของ Explainable Linear Model ต่างๆ ส่วน Proximity Measure PIE จะบอกถึงจำนวนเพื่อนบ้านที่ใกล้เคียง x ที่จะเอามาใช้ในการอธิบาย ตอน Train model เราต้องกำหนด K คือจำนวน features ยิ่ง k น้อยยิ่งง่ายต่อการอธิบาย อาจจะใช้ lasso model โดยค่อยๆเปลี่ยนค่า λ จนได้จำนวน features ที่ต้องการ หรือออีกวิธีจะทำ แบบ trained ด้วย Full model ไปเลย โดยมีทุก features แล้วค่อยเอามาทำ backward selection เอาแต่ตัวที่มี weight สูงๆ ก้ได้
  • #17: Pros สามาถใช้ complex model อะไรก้ได้โดยใช้ interpretable แบบเดิม ถ้าเราใช lasso หรือ short tree ในการ อธบายเราจะได้คำอธิบายที่ง่าย และสั้นกระชับ + human-friendly explanations LIME เป็นหนึ่งในไม่กี่อันที่สามารถใช้ในการอธบายข้อมูลที่เป็น Tabular , Text และ Image ได้ LIME package ใน python กับ R ใเรียกใช้ง่าย ในการเทรนโมเดลที่อธบายได้เราสามารถใช้ features ที่ไม่ต้องเหมืนอกับโมดลที่ซับซ้อนได้ ในกรณีที่เราต้องการทำ PCA ก่อนแล้วค่อยเอา new feature มาอธิบาย Cons การ sampling ข้อมูลต้องมีการพัฒนาเพิ่ม LIME ใช้ Gaussian dist แต่ไม่ได้คำนึงถึงการ ที่ feature มีความ correlation has to define complexity in advance Instability of explanation: in my experience, if you repeat the sampling process, then the explantions that come out can be different. Instability means that it is difficult to trust the explanations, and you should be very critical. this assumption is correct when looking at a very small region around the data sample. By expanding this region however, it is possible that a linear model might not be powerful enough to explain the behavior of the original model