SlideShare a Scribd company logo
1 of 14
Download to read offline
2021.05.21
洪 嘉源
株式会社 Mobility Technologies
Long-Tailed Classification by
Keeping the Good and Removing
the Bad Momentum Causal Effect
2
▪ この論文を選んだ理由:
▪ 現在long-tailed問題の主流手法re-sampling/re-weightingと異なる
新しい角度からlong-tailed問題の汎用手法を提案
▪ 前の発表:long-tailed classification の最新動向について[link]
▪ one-stageで手法がシンプル、尚且つ性能がいい(コードも公開済み)
▪ 手法はシンプルだが、推論過程が難解でまだ完全に理解できないため、
その部分について簡略化する
今回の論文について
3
Agenda
01|Long-tailed問題の概要
02|論文の概要
03|論文の手法
04|推論過程
05|他の手法と比較
4
▪ 各クラスをインスタンス数降順で並べて、クラスを二種類に分ける
▪ head classes: 少数のクラス、データ数が多い、学習しやすい
▪ tail classes: 多数のクラス、データ数が少ない、学習しずらい
▪ 二つのbasic approach:
▪ re-sampling: 学習データをdata balancedにサンプリングする
▪ re-weighting: ロスに重みをかけてdata balancedにする
Long-tailedなクラス分類問題へのアプローチ
head classes tail classes
cat
vulpes lagopus
5
Decoupling Representation and Classifier for Long-Tailed
Recognition, ICLR 2020 [1]
▪ 現在long-tailed classificationでSOTA(state of art)の一つ
▪ 分類パーフォマンス = 表現クオリティー + 分類器クオリティー
▪ long-tailed分布のデータを通常の学習なら分類器クオリティーが低い
▪ 直接rebalancingの手法を全モデルに適用すると表現クオリティーが下がる
▪ 1回目の学習ででlong-tailed dataをそのまま学習し、
2回目の学習で特徴抽出部の重みを固定し、データをrebalancingして
再度学習する
Long-tailed問題への代表的なアプローチ
6
Long-Tailed Classification by Keeping the Good and
Removing the Bad Momentum Causal Effect, NIPS2020
[arxiv] [github]
▪ 因果分析の角度から、学習時momentumの中のhead classesによる悪
影響を取り除くことによって、long-tailed classificationの性能を向
上する手法
▪ momentumを主役にするきっかけ:
key problemはデータではなく、学習の方法にある。optimizerのmomentumが学
習時にデータの分布を取り込んでいることに気づく、そこから着手する
今回紹介する論文の概要
7
1. de-confound training [code]
▪ 学習時multi-head normalized classifierを使うだけ
各クラスのlogits: 𝑌! =
"
#
∑$%&
# ((!
"
)#*"
(!
" +, *"
(式①),
Kはmulti-head数, 𝜏, 𝛾はhyper parameter
▪ 分母 𝑤!
$
+ 𝜆 𝑥$ を他のnormalization形式に変えてもいける
(例えば 𝑤!
$
𝑥$ )
2. 学習の同時に移動平均特徴 ̅
𝑥を統計し、そのunit vectorは特徴が
head classesへの傾向方向とみなす
▪ +
𝑑 = -
̅
*
* (式②)
De-confound-TDE手法(結論)
8
3. counterfactual TDE inference[code]
▪ 推論時下記式でlogitsを計算
各クラスのlogits: 𝑇𝐷𝐸 𝑌! =
"
#
∑$%&
#
(
((!
"
)#*"
(!
" +, *"
− 𝛼 4
./0 *", 2
3" 4((!
"
)# 2
3"
(!
" +,
) (式③),
𝛼はhyper parameter
▪ logitsからhead classesへの傾向を取り除くため
4. detection/segmentationにおいて、background classがhead class
だが処理しなくていい(non-object regionを大量消すため)[code]
▪ 推論時下記式で最後の確率を計算:
𝑎𝑟𝑔𝑚𝑎𝑥!∈6 :
1 − 𝑝7 4
𝑞!
1 − 𝑞7
𝑝7
▪ 𝑖 = 0は背景、 𝑝!は元logitsで計算した確率、 𝑞!はTDE logitsで計算した確率
De-confound-TDE手法(結論)
▪ Causal graph(因果関係図)
▪ M(momentum)はXとYのconfounder(交絡因子)
鍛錬<-年齢->癌
▪ D(特徴がhead classesへの偏移)はXとYのmediator(媒介因子)
薬->プラセボ効果->治癒
9
推論過程(よく理解できていない部分)
▪ De-confound-TDE:
▪ 最終の目的:XがYへのTDE(True Direct Effect)を求める
▪ de-confound-training(式①):MがXへの効果を抑制する
▪ counterfactual TDE inference (式③):XがYへの媒介効果を除去
▪ 厳密の数学推論ではなく、因果分析の考え方を基づいて数学言語で再定義した工程
実現
1)Mの真実分布を統計できないため、式①でmulti-head samplingで近似
2)原始のlogitsをXからYへの因果効果と見なすと、 propensity scoreの考え方を基づいて、
全てのクラスをnormalizeするが必要。logitsのnormalizationで実現
10
推論過程(よく理解できていない部分)
▪ 因果分析から見る各手法
▪ one-stage re-balancing: controlled direct effect(Dを固定値と設定)
▪ two-stage re-balancing: natural direct effect(Dの影響を完全に除去?)
▪ 各手法のサンプル
11
他の手法との比較
▪ 可視化
▪ 提案手法のfeature mapは区分度が高いregionに集中(例:warthogの牙)
12
他の手法との比較
▪ classification SOTA
▪ CIFAR-10, CIFAR-100, ImageNet-LT
▪ detection/segmentation SOTA
▪ LVIS
13
他の手法との比較
14
紹介論文:
Kaihua Tang, Jianqiang Huang, Hanwang Zhang. Long-Tailed Classification by Keeping
the Good and Removing the Bad Momentum Causal Effect. In NeurIPS, 2020.
[1] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi
Feng, Yannis Kalantidis. Decoupling Representation and Classifier for Long-Tailed
Recognition. In ICLR, 2020.
Reference

More Related Content

Similar to 論文紹介: Long-Tailed Classification by Keeping the Good and Removing the Bad Momentum Causal Effect

Kanban coaching masterclass- Ravi's notes
Kanban coaching masterclass- Ravi's notesKanban coaching masterclass- Ravi's notes
Kanban coaching masterclass- Ravi's notesRavi Tadwalkar
 
03 unified process
03 unified process03 unified process
03 unified processBaskarkncet
 
FIE 2008 Pedagogy Cybernetics
FIE 2008 Pedagogy CyberneticsFIE 2008 Pedagogy Cybernetics
FIE 2008 Pedagogy Cyberneticsarammann
 
Action research - OD process - Organizational Change and Development - Manu...
Action research  - OD process -  Organizational Change and Development - Manu...Action research  - OD process -  Organizational Change and Development - Manu...
Action research - OD process - Organizational Change and Development - Manu...manumelwin
 
TRADITIONAL AND AGILE PROJECT MANAGEMENT(KANBAN)
TRADITIONAL AND AGILE PROJECT MANAGEMENT(KANBAN)TRADITIONAL AND AGILE PROJECT MANAGEMENT(KANBAN)
TRADITIONAL AND AGILE PROJECT MANAGEMENT(KANBAN)GEORGEOFORI7
 
Project management zarkovic
Project management zarkovicProject management zarkovic
Project management zarkovicMR Z
 
Visualisation&agile practices ai2014
Visualisation&agile practices ai2014Visualisation&agile practices ai2014
Visualisation&agile practices ai2014Balaji Muniraja
 
VII Jornadas eMadrid "Education in exponential times". Erkan Er: "Predicting ...
VII Jornadas eMadrid "Education in exponential times". Erkan Er: "Predicting ...VII Jornadas eMadrid "Education in exponential times". Erkan Er: "Predicting ...
VII Jornadas eMadrid "Education in exponential times". Erkan Er: "Predicting ...eMadrid network
 
NS-CUK Seminar: J.H.Lee, Review on "Task Relation-aware Continual User Repres...
NS-CUK Seminar: J.H.Lee, Review on "Task Relation-aware Continual User Repres...NS-CUK Seminar: J.H.Lee, Review on "Task Relation-aware Continual User Repres...
NS-CUK Seminar: J.H.Lee, Review on "Task Relation-aware Continual User Repres...ssuser4b1f48
 
OO Development 2 - Software Development Methodologies
OO Development 2 - Software Development MethodologiesOO Development 2 - Software Development Methodologies
OO Development 2 - Software Development MethodologiesRandy Connolly
 
TRADITIONAL AND AGILE(KANBAN) PROJECT MANAGEMENT.
TRADITIONAL AND AGILE(KANBAN) PROJECT MANAGEMENT.TRADITIONAL AND AGILE(KANBAN) PROJECT MANAGEMENT.
TRADITIONAL AND AGILE(KANBAN) PROJECT MANAGEMENT.GEORGEOFORI7
 
TRADITIONAL AND AGILE(KANBAN) PROJECT MANAGEMENT.
TRADITIONAL AND AGILE(KANBAN) PROJECT MANAGEMENT.TRADITIONAL AND AGILE(KANBAN) PROJECT MANAGEMENT.
TRADITIONAL AND AGILE(KANBAN) PROJECT MANAGEMENT.GEORGEOFORI7
 
Alianna Maren STOMP ePortfolio
Alianna Maren STOMP ePortfolioAlianna Maren STOMP ePortfolio
Alianna Maren STOMP ePortfoliosburakharper
 
· Choose an information system for an individual project.  During .docx
· Choose an information system for an individual project.  During .docx· Choose an information system for an individual project.  During .docx
· Choose an information system for an individual project.  During .docxLynellBull52
 
Effective Software Design
Effective Software Design Effective Software Design
Effective Software Design Darshan Ashpal
 

Similar to 論文紹介: Long-Tailed Classification by Keeping the Good and Removing the Bad Momentum Causal Effect (20)

Kanban coaching masterclass- Ravi's notes
Kanban coaching masterclass- Ravi's notesKanban coaching masterclass- Ravi's notes
Kanban coaching masterclass- Ravi's notes
 
03 unified process
03 unified process03 unified process
03 unified process
 
RBT - GSN.pptx
RBT - GSN.pptxRBT - GSN.pptx
RBT - GSN.pptx
 
Usyd.tbl.dalziel.sep17
Usyd.tbl.dalziel.sep17Usyd.tbl.dalziel.sep17
Usyd.tbl.dalziel.sep17
 
Facilitation and project management, 24th Feb, Marcel Ekkel
Facilitation and project management, 24th Feb, Marcel EkkelFacilitation and project management, 24th Feb, Marcel Ekkel
Facilitation and project management, 24th Feb, Marcel Ekkel
 
FIE 2008 Pedagogy Cybernetics
FIE 2008 Pedagogy CyberneticsFIE 2008 Pedagogy Cybernetics
FIE 2008 Pedagogy Cybernetics
 
Action research - OD process - Organizational Change and Development - Manu...
Action research  - OD process -  Organizational Change and Development - Manu...Action research  - OD process -  Organizational Change and Development - Manu...
Action research - OD process - Organizational Change and Development - Manu...
 
TRADITIONAL AND AGILE PROJECT MANAGEMENT(KANBAN)
TRADITIONAL AND AGILE PROJECT MANAGEMENT(KANBAN)TRADITIONAL AND AGILE PROJECT MANAGEMENT(KANBAN)
TRADITIONAL AND AGILE PROJECT MANAGEMENT(KANBAN)
 
Project management zarkovic
Project management zarkovicProject management zarkovic
Project management zarkovic
 
Visualisation&agile practices ai2014
Visualisation&agile practices ai2014Visualisation&agile practices ai2014
Visualisation&agile practices ai2014
 
ID, UP, & RUP.pptx
ID, UP, & RUP.pptxID, UP, & RUP.pptx
ID, UP, & RUP.pptx
 
VII Jornadas eMadrid "Education in exponential times". Erkan Er: "Predicting ...
VII Jornadas eMadrid "Education in exponential times". Erkan Er: "Predicting ...VII Jornadas eMadrid "Education in exponential times". Erkan Er: "Predicting ...
VII Jornadas eMadrid "Education in exponential times". Erkan Er: "Predicting ...
 
NS-CUK Seminar: J.H.Lee, Review on "Task Relation-aware Continual User Repres...
NS-CUK Seminar: J.H.Lee, Review on "Task Relation-aware Continual User Repres...NS-CUK Seminar: J.H.Lee, Review on "Task Relation-aware Continual User Repres...
NS-CUK Seminar: J.H.Lee, Review on "Task Relation-aware Continual User Repres...
 
Purpose of research
Purpose of researchPurpose of research
Purpose of research
 
OO Development 2 - Software Development Methodologies
OO Development 2 - Software Development MethodologiesOO Development 2 - Software Development Methodologies
OO Development 2 - Software Development Methodologies
 
TRADITIONAL AND AGILE(KANBAN) PROJECT MANAGEMENT.
TRADITIONAL AND AGILE(KANBAN) PROJECT MANAGEMENT.TRADITIONAL AND AGILE(KANBAN) PROJECT MANAGEMENT.
TRADITIONAL AND AGILE(KANBAN) PROJECT MANAGEMENT.
 
TRADITIONAL AND AGILE(KANBAN) PROJECT MANAGEMENT.
TRADITIONAL AND AGILE(KANBAN) PROJECT MANAGEMENT.TRADITIONAL AND AGILE(KANBAN) PROJECT MANAGEMENT.
TRADITIONAL AND AGILE(KANBAN) PROJECT MANAGEMENT.
 
Alianna Maren STOMP ePortfolio
Alianna Maren STOMP ePortfolioAlianna Maren STOMP ePortfolio
Alianna Maren STOMP ePortfolio
 
· Choose an information system for an individual project.  During .docx
· Choose an information system for an individual project.  During .docx· Choose an information system for an individual project.  During .docx
· Choose an information system for an individual project.  During .docx
 
Effective Software Design
Effective Software Design Effective Software Design
Effective Software Design
 

More from Plot Hong

Noisy Labels と戦う深層学習
Noisy Labels と戦う深層学習Noisy Labels と戦う深層学習
Noisy Labels と戦う深層学習Plot Hong
 
PolyLoss: A POLYNOMIAL EXPANSION PERSPECTIVE OF CLASSIFICATION LOSS FUNCTION...
PolyLoss:  A POLYNOMIAL EXPANSION PERSPECTIVE OF CLASSIFICATION LOSS FUNCTION...PolyLoss:  A POLYNOMIAL EXPANSION PERSPECTIVE OF CLASSIFICATION LOSS FUNCTION...
PolyLoss: A POLYNOMIAL EXPANSION PERSPECTIVE OF CLASSIFICATION LOSS FUNCTION...Plot Hong
 
SynFace: Face Recognition with Synthetic Data 論文紹介
SynFace:  Face Recognition with Synthetic Data 論文紹介SynFace:  Face Recognition with Synthetic Data 論文紹介
SynFace: Face Recognition with Synthetic Data 論文紹介Plot Hong
 
Face Quality Assessment 顔画像品質評価について
Face Quality Assessment 顔画像品質評価についてFace Quality Assessment 顔画像品質評価について
Face Quality Assessment 顔画像品質評価についてPlot Hong
 
Long-Tailed Classificationの最新動向について
Long-Tailed Classificationの最新動向についてLong-Tailed Classificationの最新動向について
Long-Tailed Classificationの最新動向についてPlot Hong
 
Crowd Counting & Detection論文紹介
Crowd Counting & Detection論文紹介Crowd Counting & Detection論文紹介
Crowd Counting & Detection論文紹介Plot Hong
 
Deepfakesの生成および検出
Deepfakesの生成および検出Deepfakesの生成および検出
Deepfakesの生成および検出Plot Hong
 

More from Plot Hong (7)

Noisy Labels と戦う深層学習
Noisy Labels と戦う深層学習Noisy Labels と戦う深層学習
Noisy Labels と戦う深層学習
 
PolyLoss: A POLYNOMIAL EXPANSION PERSPECTIVE OF CLASSIFICATION LOSS FUNCTION...
PolyLoss:  A POLYNOMIAL EXPANSION PERSPECTIVE OF CLASSIFICATION LOSS FUNCTION...PolyLoss:  A POLYNOMIAL EXPANSION PERSPECTIVE OF CLASSIFICATION LOSS FUNCTION...
PolyLoss: A POLYNOMIAL EXPANSION PERSPECTIVE OF CLASSIFICATION LOSS FUNCTION...
 
SynFace: Face Recognition with Synthetic Data 論文紹介
SynFace:  Face Recognition with Synthetic Data 論文紹介SynFace:  Face Recognition with Synthetic Data 論文紹介
SynFace: Face Recognition with Synthetic Data 論文紹介
 
Face Quality Assessment 顔画像品質評価について
Face Quality Assessment 顔画像品質評価についてFace Quality Assessment 顔画像品質評価について
Face Quality Assessment 顔画像品質評価について
 
Long-Tailed Classificationの最新動向について
Long-Tailed Classificationの最新動向についてLong-Tailed Classificationの最新動向について
Long-Tailed Classificationの最新動向について
 
Crowd Counting & Detection論文紹介
Crowd Counting & Detection論文紹介Crowd Counting & Detection論文紹介
Crowd Counting & Detection論文紹介
 
Deepfakesの生成および検出
Deepfakesの生成および検出Deepfakesの生成および検出
Deepfakesの生成および検出
 

Recently uploaded

Install Stable Diffusion in windows machine
Install Stable Diffusion in windows machineInstall Stable Diffusion in windows machine
Install Stable Diffusion in windows machinePadma Pradeep
 
#StandardsGoals for 2024: What’s new for BISAC - Tech Forum 2024
#StandardsGoals for 2024: What’s new for BISAC - Tech Forum 2024#StandardsGoals for 2024: What’s new for BISAC - Tech Forum 2024
#StandardsGoals for 2024: What’s new for BISAC - Tech Forum 2024BookNet Canada
 
The Codex of Business Writing Software for Real-World Solutions 2.pptx
The Codex of Business Writing Software for Real-World Solutions 2.pptxThe Codex of Business Writing Software for Real-World Solutions 2.pptx
The Codex of Business Writing Software for Real-World Solutions 2.pptxMalak Abu Hammad
 
Neo4j - How KGs are shaping the future of Generative AI at AWS Summit London ...
Neo4j - How KGs are shaping the future of Generative AI at AWS Summit London ...Neo4j - How KGs are shaping the future of Generative AI at AWS Summit London ...
Neo4j - How KGs are shaping the future of Generative AI at AWS Summit London ...Neo4j
 
My Hashitalk Indonesia April 2024 Presentation
My Hashitalk Indonesia April 2024 PresentationMy Hashitalk Indonesia April 2024 Presentation
My Hashitalk Indonesia April 2024 PresentationRidwan Fadjar
 
IAC 2024 - IA Fast Track to Search Focused AI Solutions
IAC 2024 - IA Fast Track to Search Focused AI SolutionsIAC 2024 - IA Fast Track to Search Focused AI Solutions
IAC 2024 - IA Fast Track to Search Focused AI SolutionsEnterprise Knowledge
 
Maximizing Board Effectiveness 2024 Webinar.pptx
Maximizing Board Effectiveness 2024 Webinar.pptxMaximizing Board Effectiveness 2024 Webinar.pptx
Maximizing Board Effectiveness 2024 Webinar.pptxOnBoard
 
FULL ENJOY 🔝 8264348440 🔝 Call Girls in Diplomatic Enclave | Delhi
FULL ENJOY 🔝 8264348440 🔝 Call Girls in Diplomatic Enclave | DelhiFULL ENJOY 🔝 8264348440 🔝 Call Girls in Diplomatic Enclave | Delhi
FULL ENJOY 🔝 8264348440 🔝 Call Girls in Diplomatic Enclave | Delhisoniya singh
 
Azure Monitor & Application Insight to monitor Infrastructure & Application
Azure Monitor & Application Insight to monitor Infrastructure & ApplicationAzure Monitor & Application Insight to monitor Infrastructure & Application
Azure Monitor & Application Insight to monitor Infrastructure & ApplicationAndikSusilo4
 
Snow Chain-Integrated Tire for a Safe Drive on Winter Roads
Snow Chain-Integrated Tire for a Safe Drive on Winter RoadsSnow Chain-Integrated Tire for a Safe Drive on Winter Roads
Snow Chain-Integrated Tire for a Safe Drive on Winter RoadsHyundai Motor Group
 
Next-generation AAM aircraft unveiled by Supernal, S-A2
Next-generation AAM aircraft unveiled by Supernal, S-A2Next-generation AAM aircraft unveiled by Supernal, S-A2
Next-generation AAM aircraft unveiled by Supernal, S-A2Hyundai Motor Group
 
Making_way_through_DLL_hollowing_inspite_of_CFG_by_Debjeet Banerjee.pptx
Making_way_through_DLL_hollowing_inspite_of_CFG_by_Debjeet Banerjee.pptxMaking_way_through_DLL_hollowing_inspite_of_CFG_by_Debjeet Banerjee.pptx
Making_way_through_DLL_hollowing_inspite_of_CFG_by_Debjeet Banerjee.pptxnull - The Open Security Community
 
Beyond Boundaries: Leveraging No-Code Solutions for Industry Innovation
Beyond Boundaries: Leveraging No-Code Solutions for Industry InnovationBeyond Boundaries: Leveraging No-Code Solutions for Industry Innovation
Beyond Boundaries: Leveraging No-Code Solutions for Industry InnovationSafe Software
 
08448380779 Call Girls In Greater Kailash - I Women Seeking Men
08448380779 Call Girls In Greater Kailash - I Women Seeking Men08448380779 Call Girls In Greater Kailash - I Women Seeking Men
08448380779 Call Girls In Greater Kailash - I Women Seeking MenDelhi Call girls
 
Transcript: #StandardsGoals for 2024: What’s new for BISAC - Tech Forum 2024
Transcript: #StandardsGoals for 2024: What’s new for BISAC - Tech Forum 2024Transcript: #StandardsGoals for 2024: What’s new for BISAC - Tech Forum 2024
Transcript: #StandardsGoals for 2024: What’s new for BISAC - Tech Forum 2024BookNet Canada
 
Integration and Automation in Practice: CI/CD in Mule Integration and Automat...
Integration and Automation in Practice: CI/CD in Mule Integration and Automat...Integration and Automation in Practice: CI/CD in Mule Integration and Automat...
Integration and Automation in Practice: CI/CD in Mule Integration and Automat...Patryk Bandurski
 
Injustice - Developers Among Us (SciFiDevCon 2024)
Injustice - Developers Among Us (SciFiDevCon 2024)Injustice - Developers Among Us (SciFiDevCon 2024)
Injustice - Developers Among Us (SciFiDevCon 2024)Allon Mureinik
 

Recently uploaded (20)

Install Stable Diffusion in windows machine
Install Stable Diffusion in windows machineInstall Stable Diffusion in windows machine
Install Stable Diffusion in windows machine
 
E-Vehicle_Hacking_by_Parul Sharma_null_owasp.pptx
E-Vehicle_Hacking_by_Parul Sharma_null_owasp.pptxE-Vehicle_Hacking_by_Parul Sharma_null_owasp.pptx
E-Vehicle_Hacking_by_Parul Sharma_null_owasp.pptx
 
The transition to renewables in India.pdf
The transition to renewables in India.pdfThe transition to renewables in India.pdf
The transition to renewables in India.pdf
 
#StandardsGoals for 2024: What’s new for BISAC - Tech Forum 2024
#StandardsGoals for 2024: What’s new for BISAC - Tech Forum 2024#StandardsGoals for 2024: What’s new for BISAC - Tech Forum 2024
#StandardsGoals for 2024: What’s new for BISAC - Tech Forum 2024
 
The Codex of Business Writing Software for Real-World Solutions 2.pptx
The Codex of Business Writing Software for Real-World Solutions 2.pptxThe Codex of Business Writing Software for Real-World Solutions 2.pptx
The Codex of Business Writing Software for Real-World Solutions 2.pptx
 
Neo4j - How KGs are shaping the future of Generative AI at AWS Summit London ...
Neo4j - How KGs are shaping the future of Generative AI at AWS Summit London ...Neo4j - How KGs are shaping the future of Generative AI at AWS Summit London ...
Neo4j - How KGs are shaping the future of Generative AI at AWS Summit London ...
 
My Hashitalk Indonesia April 2024 Presentation
My Hashitalk Indonesia April 2024 PresentationMy Hashitalk Indonesia April 2024 Presentation
My Hashitalk Indonesia April 2024 Presentation
 
IAC 2024 - IA Fast Track to Search Focused AI Solutions
IAC 2024 - IA Fast Track to Search Focused AI SolutionsIAC 2024 - IA Fast Track to Search Focused AI Solutions
IAC 2024 - IA Fast Track to Search Focused AI Solutions
 
Maximizing Board Effectiveness 2024 Webinar.pptx
Maximizing Board Effectiveness 2024 Webinar.pptxMaximizing Board Effectiveness 2024 Webinar.pptx
Maximizing Board Effectiveness 2024 Webinar.pptx
 
FULL ENJOY 🔝 8264348440 🔝 Call Girls in Diplomatic Enclave | Delhi
FULL ENJOY 🔝 8264348440 🔝 Call Girls in Diplomatic Enclave | DelhiFULL ENJOY 🔝 8264348440 🔝 Call Girls in Diplomatic Enclave | Delhi
FULL ENJOY 🔝 8264348440 🔝 Call Girls in Diplomatic Enclave | Delhi
 
Azure Monitor & Application Insight to monitor Infrastructure & Application
Azure Monitor & Application Insight to monitor Infrastructure & ApplicationAzure Monitor & Application Insight to monitor Infrastructure & Application
Azure Monitor & Application Insight to monitor Infrastructure & Application
 
Snow Chain-Integrated Tire for a Safe Drive on Winter Roads
Snow Chain-Integrated Tire for a Safe Drive on Winter RoadsSnow Chain-Integrated Tire for a Safe Drive on Winter Roads
Snow Chain-Integrated Tire for a Safe Drive on Winter Roads
 
Next-generation AAM aircraft unveiled by Supernal, S-A2
Next-generation AAM aircraft unveiled by Supernal, S-A2Next-generation AAM aircraft unveiled by Supernal, S-A2
Next-generation AAM aircraft unveiled by Supernal, S-A2
 
Making_way_through_DLL_hollowing_inspite_of_CFG_by_Debjeet Banerjee.pptx
Making_way_through_DLL_hollowing_inspite_of_CFG_by_Debjeet Banerjee.pptxMaking_way_through_DLL_hollowing_inspite_of_CFG_by_Debjeet Banerjee.pptx
Making_way_through_DLL_hollowing_inspite_of_CFG_by_Debjeet Banerjee.pptx
 
Vulnerability_Management_GRC_by Sohang Sengupta.pptx
Vulnerability_Management_GRC_by Sohang Sengupta.pptxVulnerability_Management_GRC_by Sohang Sengupta.pptx
Vulnerability_Management_GRC_by Sohang Sengupta.pptx
 
Beyond Boundaries: Leveraging No-Code Solutions for Industry Innovation
Beyond Boundaries: Leveraging No-Code Solutions for Industry InnovationBeyond Boundaries: Leveraging No-Code Solutions for Industry Innovation
Beyond Boundaries: Leveraging No-Code Solutions for Industry Innovation
 
08448380779 Call Girls In Greater Kailash - I Women Seeking Men
08448380779 Call Girls In Greater Kailash - I Women Seeking Men08448380779 Call Girls In Greater Kailash - I Women Seeking Men
08448380779 Call Girls In Greater Kailash - I Women Seeking Men
 
Transcript: #StandardsGoals for 2024: What’s new for BISAC - Tech Forum 2024
Transcript: #StandardsGoals for 2024: What’s new for BISAC - Tech Forum 2024Transcript: #StandardsGoals for 2024: What’s new for BISAC - Tech Forum 2024
Transcript: #StandardsGoals for 2024: What’s new for BISAC - Tech Forum 2024
 
Integration and Automation in Practice: CI/CD in Mule Integration and Automat...
Integration and Automation in Practice: CI/CD in Mule Integration and Automat...Integration and Automation in Practice: CI/CD in Mule Integration and Automat...
Integration and Automation in Practice: CI/CD in Mule Integration and Automat...
 
Injustice - Developers Among Us (SciFiDevCon 2024)
Injustice - Developers Among Us (SciFiDevCon 2024)Injustice - Developers Among Us (SciFiDevCon 2024)
Injustice - Developers Among Us (SciFiDevCon 2024)
 

論文紹介: Long-Tailed Classification by Keeping the Good and Removing the Bad Momentum Causal Effect

  • 1. 2021.05.21 洪 嘉源 株式会社 Mobility Technologies Long-Tailed Classification by Keeping the Good and Removing the Bad Momentum Causal Effect
  • 2. 2 ▪ この論文を選んだ理由: ▪ 現在long-tailed問題の主流手法re-sampling/re-weightingと異なる 新しい角度からlong-tailed問題の汎用手法を提案 ▪ 前の発表:long-tailed classification の最新動向について[link] ▪ one-stageで手法がシンプル、尚且つ性能がいい(コードも公開済み) ▪ 手法はシンプルだが、推論過程が難解でまだ完全に理解できないため、 その部分について簡略化する 今回の論文について
  • 4. 4 ▪ 各クラスをインスタンス数降順で並べて、クラスを二種類に分ける ▪ head classes: 少数のクラス、データ数が多い、学習しやすい ▪ tail classes: 多数のクラス、データ数が少ない、学習しずらい ▪ 二つのbasic approach: ▪ re-sampling: 学習データをdata balancedにサンプリングする ▪ re-weighting: ロスに重みをかけてdata balancedにする Long-tailedなクラス分類問題へのアプローチ head classes tail classes cat vulpes lagopus
  • 5. 5 Decoupling Representation and Classifier for Long-Tailed Recognition, ICLR 2020 [1] ▪ 現在long-tailed classificationでSOTA(state of art)の一つ ▪ 分類パーフォマンス = 表現クオリティー + 分類器クオリティー ▪ long-tailed分布のデータを通常の学習なら分類器クオリティーが低い ▪ 直接rebalancingの手法を全モデルに適用すると表現クオリティーが下がる ▪ 1回目の学習ででlong-tailed dataをそのまま学習し、 2回目の学習で特徴抽出部の重みを固定し、データをrebalancingして 再度学習する Long-tailed問題への代表的なアプローチ
  • 6. 6 Long-Tailed Classification by Keeping the Good and Removing the Bad Momentum Causal Effect, NIPS2020 [arxiv] [github] ▪ 因果分析の角度から、学習時momentumの中のhead classesによる悪 影響を取り除くことによって、long-tailed classificationの性能を向 上する手法 ▪ momentumを主役にするきっかけ: key problemはデータではなく、学習の方法にある。optimizerのmomentumが学 習時にデータの分布を取り込んでいることに気づく、そこから着手する 今回紹介する論文の概要
  • 7. 7 1. de-confound training [code] ▪ 学習時multi-head normalized classifierを使うだけ 各クラスのlogits: 𝑌! = " # ∑$%& # ((! " )#*" (! " +, *" (式①), Kはmulti-head数, 𝜏, 𝛾はhyper parameter ▪ 分母 𝑤! $ + 𝜆 𝑥$ を他のnormalization形式に変えてもいける (例えば 𝑤! $ 𝑥$ ) 2. 学習の同時に移動平均特徴 ̅ 𝑥を統計し、そのunit vectorは特徴が head classesへの傾向方向とみなす ▪ + 𝑑 = - ̅ * * (式②) De-confound-TDE手法(結論)
  • 8. 8 3. counterfactual TDE inference[code] ▪ 推論時下記式でlogitsを計算 各クラスのlogits: 𝑇𝐷𝐸 𝑌! = " # ∑$%& # ( ((! " )#*" (! " +, *" − 𝛼 4 ./0 *", 2 3" 4((! " )# 2 3" (! " +, ) (式③), 𝛼はhyper parameter ▪ logitsからhead classesへの傾向を取り除くため 4. detection/segmentationにおいて、background classがhead class だが処理しなくていい(non-object regionを大量消すため)[code] ▪ 推論時下記式で最後の確率を計算: 𝑎𝑟𝑔𝑚𝑎𝑥!∈6 : 1 − 𝑝7 4 𝑞! 1 − 𝑞7 𝑝7 ▪ 𝑖 = 0は背景、 𝑝!は元logitsで計算した確率、 𝑞!はTDE logitsで計算した確率 De-confound-TDE手法(結論)
  • 9. ▪ Causal graph(因果関係図) ▪ M(momentum)はXとYのconfounder(交絡因子) 鍛錬<-年齢->癌 ▪ D(特徴がhead classesへの偏移)はXとYのmediator(媒介因子) 薬->プラセボ効果->治癒 9 推論過程(よく理解できていない部分)
  • 10. ▪ De-confound-TDE: ▪ 最終の目的:XがYへのTDE(True Direct Effect)を求める ▪ de-confound-training(式①):MがXへの効果を抑制する ▪ counterfactual TDE inference (式③):XがYへの媒介効果を除去 ▪ 厳密の数学推論ではなく、因果分析の考え方を基づいて数学言語で再定義した工程 実現 1)Mの真実分布を統計できないため、式①でmulti-head samplingで近似 2)原始のlogitsをXからYへの因果効果と見なすと、 propensity scoreの考え方を基づいて、 全てのクラスをnormalizeするが必要。logitsのnormalizationで実現 10 推論過程(よく理解できていない部分)
  • 11. ▪ 因果分析から見る各手法 ▪ one-stage re-balancing: controlled direct effect(Dを固定値と設定) ▪ two-stage re-balancing: natural direct effect(Dの影響を完全に除去?) ▪ 各手法のサンプル 11 他の手法との比較
  • 12. ▪ 可視化 ▪ 提案手法のfeature mapは区分度が高いregionに集中(例:warthogの牙) 12 他の手法との比較
  • 13. ▪ classification SOTA ▪ CIFAR-10, CIFAR-100, ImageNet-LT ▪ detection/segmentation SOTA ▪ LVIS 13 他の手法との比較
  • 14. 14 紹介論文: Kaihua Tang, Jianqiang Huang, Hanwang Zhang. Long-Tailed Classification by Keeping the Good and Removing the Bad Momentum Causal Effect. In NeurIPS, 2020. [1] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, Yannis Kalantidis. Decoupling Representation and Classifier for Long-Tailed Recognition. In ICLR, 2020. Reference