SlideShare una empresa de Scribd logo
1 de 13
Masayuki Tanaka
Breaking Inter-Layer Co-Adaptation
by Classifier Anonymization
Ikuro Sato†, Kohta Ishikawa†, Guoqing Liu†, Masayuki Tanaka‡
(ICML2019)
† ‡
Meta reviewer’s comment
…This paper seems to me like a perfect example of a
“High Risk High Reward” paper, …
Acceptance ratio of ICML2019: 773/3424 = 22.6%
We have taken that as a compliment. It is a research!
1
What I’m going to talk
𝑥𝑥
Input
𝐹𝐹𝜙𝜙(𝑥𝑥) 𝐶𝐶𝜃𝜃(𝜉𝜉)
𝜂𝜂
Output
𝜉𝜉
Feature
Let’s consider a classification task.
Feature extractor Classifier
+
-
Feature space 𝜉𝜉
+
+
+ +
+
+ +
--
-
-
-- -
-
Feature space 𝜉𝜉
+
++
+
+
+
+-- --
--
-
End-to-end DNN
<<
Which is better? Why? How can we obtain good features?2
Summary
About what?
How?
Theory?
In reality?
Breaking co-adaptation between
feature extractor and classifier.
By classifier anonymization technique.
Proved: Features form simple
point-like distribution.
Point-like property largely confirmed
on real datasets.
3
What is a co-adaptation?
𝑥𝑥
Input
𝐹𝐹𝜙𝜙(𝑥𝑥) 𝐶𝐶𝜃𝜃(𝜉𝜉)
𝜂𝜂
Output
𝜉𝜉
Feature
Let’s consider a classification task.
Feature extractor Classifier
+
-
Feature space 𝜉𝜉
Decision
boundary
+
+
+ +
+
+ +
--
-
-
-- -
Co-adaptation:
Feature extractor adapts a particular classifier.
Classifier adapts a particular feature extractor.
Break
co-adaptation
-
Feature space 𝜉𝜉
+
++
+
+
+
+-- --
--
-
Classifiers
Feature extractor should be
trained for many classifiers.
End-to-end DNN
4
Proposed algorithm: FOCA
-
Feature space 𝜉𝜉
+++
+
+ ++
--
-----
(Under several conditions,)
we theoretically proved the FOCA
can train the feature extractor
which projects single point.
for given feature extractor
FOCA can train feature extractor to make any weak classifier strong.
FOCA:
Feature-extractor Optimization through Classifier Anonymization
5
Message of FOCA
Traditional training FOCA training
Feature extractor
(Junior researcher)
Feature extractor
(Junior researcher)
Weak classifiers
(Boss variety???)
Strong classifier
(Smart boss)
Transfer learning
(New boss, new domain)
FOCA can train
feature extractor strong.
6
Weak classifier assumption
Definition:
Weak classifier is slightly better than random guess.
𝜃𝜃𝜙𝜙
∗
= arg min
𝜃𝜃
E
(𝑥𝑥,𝑡𝑡)~𝑝𝑝(𝑥𝑥,𝑡𝑡)
𝐿𝐿 𝐶𝐶𝜃𝜃 𝐹𝐹𝜙𝜙(𝑥𝑥) , 𝑡𝑡
𝜃𝜃𝜙𝜙
𝐵𝐵
= arg min
𝜃𝜃
�
𝑥𝑥,𝑡𝑡 ∈𝐵𝐵
𝐿𝐿 𝐶𝐶𝜃𝜃 𝐹𝐹𝜙𝜙(𝑥𝑥) , 𝑡𝑡
Strong classifier
Strong classifier is strong for entire data.
Weak classifier assumption
We assume that strong classifier for small samples is
weak classifier for entire data.
B is small samples of entire data.
7
Practical FOCA algorithm
𝐹𝐹𝜙𝜙(𝑥𝑥)
𝐶𝐶𝜃𝜃(𝜉𝜉)
Weak classifier
generatorFeature
extractor
Classifier model
𝐹𝐹𝐹𝜙𝜙(𝑥𝑥)
Previous
feature extractor
Training data
Optimize the classifier
for given small samples
with previous feature extractor.
Update feature extractor
for given mini-batch
with weak classifier.
Sampling
𝐶𝐶𝜃𝜃(𝜉𝜉)
Weak classifier
Update
Mini-batch
8
Experimental validation
Two-step training:
Train the feature extractor. Then, train the classifier with the fixed
given feature extractor.
-
Feature space 𝜉𝜉
+
+
+ +
+
+ +
--
-
-
-- -
Co-adaptation Point-like
-
Feature space 𝜉𝜉
+++
+
+ ++
--
-----
Many samples are required to train
the classifier.
A few samples are good enough to
train the classifier.
9
Results
10
Poster as a summary
11
Links
Official proceedings of ICML2019
http://proceedings.mlr.press/v97/
arxiv: Breaking Inter-Layer Co-Adaptation by Classifier Anonymization
https://arxiv.org/abs/1906.01150
Twitter: Masayuki Tanaka
https://twitter.com/likesilkto
Twitter: Ikuro Sato
https://twitter.com/ikuro_s
12

Más contenido relacionado

La actualidad más candente

La actualidad más candente (16)

Fuzzy logic member functions
Fuzzy logic member functionsFuzzy logic member functions
Fuzzy logic member functions
 
Generics
GenericsGenerics
Generics
 
Optimal feature selection from v mware esxi 5.1 feature set
Optimal feature selection from v mware esxi 5.1 feature setOptimal feature selection from v mware esxi 5.1 feature set
Optimal feature selection from v mware esxi 5.1 feature set
 
Best practices in Java
Best practices in JavaBest practices in Java
Best practices in Java
 
Wrapper classes
Wrapper classesWrapper classes
Wrapper classes
 
DotNet programming & Practices
DotNet programming & PracticesDotNet programming & Practices
DotNet programming & Practices
 
(Recursion)ads
(Recursion)ads(Recursion)ads
(Recursion)ads
 
Recursion Pattern Analysis and Feedback
Recursion Pattern Analysis and FeedbackRecursion Pattern Analysis and Feedback
Recursion Pattern Analysis and Feedback
 
Pattern Matching - at a glance
Pattern Matching - at a glancePattern Matching - at a glance
Pattern Matching - at a glance
 
Chapter 11 ds
Chapter 11 dsChapter 11 ds
Chapter 11 ds
 
Java Generics
Java GenericsJava Generics
Java Generics
 
wrapper classes
wrapper classeswrapper classes
wrapper classes
 
Feature selection
Feature selectionFeature selection
Feature selection
 
Generics in java
Generics in javaGenerics in java
Generics in java
 
Feature recognition and classification
Feature recognition and classificationFeature recognition and classification
Feature recognition and classification
 
Data Handling and Function
Data Handling and FunctionData Handling and Function
Data Handling and Function
 

Similar a Slideshare breaking inter layer co-adaptation

Reconstruction of a Complete Dataset from an Incomplete Dataset by ARA (Attri...
Reconstruction of a Complete Dataset from an Incomplete Dataset by ARA (Attri...Reconstruction of a Complete Dataset from an Incomplete Dataset by ARA (Attri...
Reconstruction of a Complete Dataset from an Incomplete Dataset by ARA (Attri...
Waqas Tariq
 
Using CNTK's Python Interface for Deep LearningDave DeBarr -
Using CNTK's Python Interface for Deep LearningDave DeBarr - Using CNTK's Python Interface for Deep LearningDave DeBarr -
Using CNTK's Python Interface for Deep LearningDave DeBarr -
PyData
 
Efficient top-k queries processing in column-family distributed databases
Efficient top-k queries processing in column-family distributed databasesEfficient top-k queries processing in column-family distributed databases
Efficient top-k queries processing in column-family distributed databases
Rui Vieira
 
XGBoostLSS - An extension of XGBoost to probabilistic forecasting, Alexander ...
XGBoostLSS - An extension of XGBoost to probabilistic forecasting, Alexander ...XGBoostLSS - An extension of XGBoost to probabilistic forecasting, Alexander ...
XGBoostLSS - An extension of XGBoost to probabilistic forecasting, Alexander ...
Erlangen Artificial Intelligence & Machine Learning Meetup
 

Similar a Slideshare breaking inter layer co-adaptation (20)

Machine learning for document analysis and understanding
Machine learning for document analysis and understandingMachine learning for document analysis and understanding
Machine learning for document analysis and understanding
 
Machine Learning Lecture 3 Decision Trees
Machine Learning Lecture 3 Decision TreesMachine Learning Lecture 3 Decision Trees
Machine Learning Lecture 3 Decision Trees
 
Reconstruction of a Complete Dataset from an Incomplete Dataset by ARA (Attri...
Reconstruction of a Complete Dataset from an Incomplete Dataset by ARA (Attri...Reconstruction of a Complete Dataset from an Incomplete Dataset by ARA (Attri...
Reconstruction of a Complete Dataset from an Incomplete Dataset by ARA (Attri...
 
Using CNTK's Python Interface for Deep LearningDave DeBarr -
Using CNTK's Python Interface for Deep LearningDave DeBarr - Using CNTK's Python Interface for Deep LearningDave DeBarr -
Using CNTK's Python Interface for Deep LearningDave DeBarr -
 
Spark Meetup
Spark MeetupSpark Meetup
Spark Meetup
 
Parameterizing and Assembling IR-based Solutions for SE Tasks using Genetic A...
Parameterizing and Assembling IR-based Solutions for SE Tasks using Genetic A...Parameterizing and Assembling IR-based Solutions for SE Tasks using Genetic A...
Parameterizing and Assembling IR-based Solutions for SE Tasks using Genetic A...
 
Text analysis using python
Text analysis using pythonText analysis using python
Text analysis using python
 
Branch And Bound and Beam Search Feature Selection Algorithms
Branch And Bound and Beam Search Feature Selection AlgorithmsBranch And Bound and Beam Search Feature Selection Algorithms
Branch And Bound and Beam Search Feature Selection Algorithms
 
supervised.pptx
supervised.pptxsupervised.pptx
supervised.pptx
 
Efficient top-k queries processing in column-family distributed databases
Efficient top-k queries processing in column-family distributed databasesEfficient top-k queries processing in column-family distributed databases
Efficient top-k queries processing in column-family distributed databases
 
Kaggle Otto Challenge: How we achieved 85th out of 3,514 and what we learnt
Kaggle Otto Challenge: How we achieved 85th out of 3,514 and what we learntKaggle Otto Challenge: How we achieved 85th out of 3,514 and what we learnt
Kaggle Otto Challenge: How we achieved 85th out of 3,514 and what we learnt
 
XGBoostLSS - An extension of XGBoost to probabilistic forecasting, Alexander ...
XGBoostLSS - An extension of XGBoost to probabilistic forecasting, Alexander ...XGBoostLSS - An extension of XGBoost to probabilistic forecasting, Alexander ...
XGBoostLSS - An extension of XGBoost to probabilistic forecasting, Alexander ...
 
Booting into functional programming
Booting into functional programmingBooting into functional programming
Booting into functional programming
 
Deep learning from a novice perspective
Deep learning from a novice perspectiveDeep learning from a novice perspective
Deep learning from a novice perspective
 
Python master class 2
Python master class 2Python master class 2
Python master class 2
 
Learning to Balance: Bayesian Meta-Learning for Imbalanced and Out-of-distrib...
Learning to Balance: Bayesian Meta-Learning for Imbalanced and Out-of-distrib...Learning to Balance: Bayesian Meta-Learning for Imbalanced and Out-of-distrib...
Learning to Balance: Bayesian Meta-Learning for Imbalanced and Out-of-distrib...
 
Foundations: Artificial Neural Networks
Foundations: Artificial Neural NetworksFoundations: Artificial Neural Networks
Foundations: Artificial Neural Networks
 
2017 nov reflow sbtb
2017 nov reflow sbtb2017 nov reflow sbtb
2017 nov reflow sbtb
 
Text Classification in Python – using Pandas, scikit-learn, IPython Notebook ...
Text Classification in Python – using Pandas, scikit-learn, IPython Notebook ...Text Classification in Python – using Pandas, scikit-learn, IPython Notebook ...
Text Classification in Python – using Pandas, scikit-learn, IPython Notebook ...
 
Getting Started with Keras and TensorFlow - StampedeCon AI Summit 2017
Getting Started with Keras and TensorFlow - StampedeCon AI Summit 2017Getting Started with Keras and TensorFlow - StampedeCon AI Summit 2017
Getting Started with Keras and TensorFlow - StampedeCon AI Summit 2017
 

Más de Masayuki Tanaka

遠赤外線カメラと可視カメラを利用した悪条件下における画像取得
遠赤外線カメラと可視カメラを利用した悪条件下における画像取得遠赤外線カメラと可視カメラを利用した悪条件下における画像取得
遠赤外線カメラと可視カメラを利用した悪条件下における画像取得
Masayuki Tanaka
 

Más de Masayuki Tanaka (20)

PRMU201902 Presentation document
PRMU201902 Presentation documentPRMU201902 Presentation document
PRMU201902 Presentation document
 
Gradient-Based Low-Light Image Enhancement
Gradient-Based Low-Light Image EnhancementGradient-Based Low-Light Image Enhancement
Gradient-Based Low-Light Image Enhancement
 
Year-End Seminar 2018
Year-End Seminar 2018Year-End Seminar 2018
Year-End Seminar 2018
 
遠赤外線カメラと可視カメラを利用した悪条件下における画像取得
遠赤外線カメラと可視カメラを利用した悪条件下における画像取得遠赤外線カメラと可視カメラを利用した悪条件下における画像取得
遠赤外線カメラと可視カメラを利用した悪条件下における画像取得
 
Learnable Image Encryption
Learnable Image EncryptionLearnable Image Encryption
Learnable Image Encryption
 
クリエイティブ・コモンズ
クリエイティブ・コモンズクリエイティブ・コモンズ
クリエイティブ・コモンズ
 
デザイン4原則
デザイン4原則デザイン4原則
デザイン4原則
 
メラビアンの法則
メラビアンの法則メラビアンの法則
メラビアンの法則
 
類似性の法則
類似性の法則類似性の法則
類似性の法則
 
権威に訴える論証
権威に訴える論証権威に訴える論証
権威に訴える論証
 
Chain rule of deep neural network layer for back propagation
Chain rule of deep neural network layer for back propagationChain rule of deep neural network layer for back propagation
Chain rule of deep neural network layer for back propagation
 
Give Me Four
Give Me FourGive Me Four
Give Me Four
 
Tech art 20170315
Tech art 20170315Tech art 20170315
Tech art 20170315
 
My Slide Theme
My Slide ThemeMy Slide Theme
My Slide Theme
 
Font Memo
Font MemoFont Memo
Font Memo
 
One-point for presentation
One-point for presentationOne-point for presentation
One-point for presentation
 
ADMM algorithm in ProxImaL
ADMM algorithm in ProxImaL ADMM algorithm in ProxImaL
ADMM algorithm in ProxImaL
 
Intensity Constraint Gradient-Based Image Reconstruction
Intensity Constraint Gradient-Based Image ReconstructionIntensity Constraint Gradient-Based Image Reconstruction
Intensity Constraint Gradient-Based Image Reconstruction
 
Least Square with L0, L1, and L2 Constraint
Least Square with L0, L1, and L2 ConstraintLeast Square with L0, L1, and L2 Constraint
Least Square with L0, L1, and L2 Constraint
 
Lasso regression
Lasso regressionLasso regression
Lasso regression
 

Último

Pests of mustard_Identification_Management_Dr.UPR.pdf
Pests of mustard_Identification_Management_Dr.UPR.pdfPests of mustard_Identification_Management_Dr.UPR.pdf
Pests of mustard_Identification_Management_Dr.UPR.pdf
PirithiRaju
 
Conjugation, transduction and transformation
Conjugation, transduction and transformationConjugation, transduction and transformation
Conjugation, transduction and transformation
Areesha Ahmad
 
Asymmetry in the atmosphere of the ultra-hot Jupiter WASP-76 b
Asymmetry in the atmosphere of the ultra-hot Jupiter WASP-76 bAsymmetry in the atmosphere of the ultra-hot Jupiter WASP-76 b
Asymmetry in the atmosphere of the ultra-hot Jupiter WASP-76 b
Sérgio Sacani
 
Pests of cotton_Sucking_Pests_Dr.UPR.pdf
Pests of cotton_Sucking_Pests_Dr.UPR.pdfPests of cotton_Sucking_Pests_Dr.UPR.pdf
Pests of cotton_Sucking_Pests_Dr.UPR.pdf
PirithiRaju
 

Último (20)

GBSN - Biochemistry (Unit 1)
GBSN - Biochemistry (Unit 1)GBSN - Biochemistry (Unit 1)
GBSN - Biochemistry (Unit 1)
 
Pulmonary drug delivery system M.pharm -2nd sem P'ceutics
Pulmonary drug delivery system M.pharm -2nd sem P'ceuticsPulmonary drug delivery system M.pharm -2nd sem P'ceutics
Pulmonary drug delivery system M.pharm -2nd sem P'ceutics
 
Pests of mustard_Identification_Management_Dr.UPR.pdf
Pests of mustard_Identification_Management_Dr.UPR.pdfPests of mustard_Identification_Management_Dr.UPR.pdf
Pests of mustard_Identification_Management_Dr.UPR.pdf
 
GBSN - Microbiology (Unit 3)
GBSN - Microbiology (Unit 3)GBSN - Microbiology (Unit 3)
GBSN - Microbiology (Unit 3)
 
Forensic Biology & Its biological significance.pdf
Forensic Biology & Its biological significance.pdfForensic Biology & Its biological significance.pdf
Forensic Biology & Its biological significance.pdf
 
Site Acceptance Test .
Site Acceptance Test                    .Site Acceptance Test                    .
Site Acceptance Test .
 
Conjugation, transduction and transformation
Conjugation, transduction and transformationConjugation, transduction and transformation
Conjugation, transduction and transformation
 
Asymmetry in the atmosphere of the ultra-hot Jupiter WASP-76 b
Asymmetry in the atmosphere of the ultra-hot Jupiter WASP-76 bAsymmetry in the atmosphere of the ultra-hot Jupiter WASP-76 b
Asymmetry in the atmosphere of the ultra-hot Jupiter WASP-76 b
 
Sector 62, Noida Call girls :8448380779 Model Escorts | 100% verified
Sector 62, Noida Call girls :8448380779 Model Escorts | 100% verifiedSector 62, Noida Call girls :8448380779 Model Escorts | 100% verified
Sector 62, Noida Call girls :8448380779 Model Escorts | 100% verified
 
chemical bonding Essentials of Physical Chemistry2.pdf
chemical bonding Essentials of Physical Chemistry2.pdfchemical bonding Essentials of Physical Chemistry2.pdf
chemical bonding Essentials of Physical Chemistry2.pdf
 
Vip profile Call Girls In Lonavala 9748763073 For Genuine Sex Service At Just...
Vip profile Call Girls In Lonavala 9748763073 For Genuine Sex Service At Just...Vip profile Call Girls In Lonavala 9748763073 For Genuine Sex Service At Just...
Vip profile Call Girls In Lonavala 9748763073 For Genuine Sex Service At Just...
 
9999266834 Call Girls In Noida Sector 22 (Delhi) Call Girl Service
9999266834 Call Girls In Noida Sector 22 (Delhi) Call Girl Service9999266834 Call Girls In Noida Sector 22 (Delhi) Call Girl Service
9999266834 Call Girls In Noida Sector 22 (Delhi) Call Girl Service
 
GBSN - Microbiology (Unit 2)
GBSN - Microbiology (Unit 2)GBSN - Microbiology (Unit 2)
GBSN - Microbiology (Unit 2)
 
SAMASTIPUR CALL GIRL 7857803690 LOW PRICE ESCORT SERVICE
SAMASTIPUR CALL GIRL 7857803690  LOW PRICE  ESCORT SERVICESAMASTIPUR CALL GIRL 7857803690  LOW PRICE  ESCORT SERVICE
SAMASTIPUR CALL GIRL 7857803690 LOW PRICE ESCORT SERVICE
 
Pests of cotton_Sucking_Pests_Dr.UPR.pdf
Pests of cotton_Sucking_Pests_Dr.UPR.pdfPests of cotton_Sucking_Pests_Dr.UPR.pdf
Pests of cotton_Sucking_Pests_Dr.UPR.pdf
 
Call Girls Ahmedabad +917728919243 call me Independent Escort Service
Call Girls Ahmedabad +917728919243 call me Independent Escort ServiceCall Girls Ahmedabad +917728919243 call me Independent Escort Service
Call Girls Ahmedabad +917728919243 call me Independent Escort Service
 
PSYCHOSOCIAL NEEDS. in nursing II sem pptx
PSYCHOSOCIAL NEEDS. in nursing II sem pptxPSYCHOSOCIAL NEEDS. in nursing II sem pptx
PSYCHOSOCIAL NEEDS. in nursing II sem pptx
 
FAIRSpectra - Enabling the FAIRification of Spectroscopy and Spectrometry
FAIRSpectra - Enabling the FAIRification of Spectroscopy and SpectrometryFAIRSpectra - Enabling the FAIRification of Spectroscopy and Spectrometry
FAIRSpectra - Enabling the FAIRification of Spectroscopy and Spectrometry
 
Dubai Call Girls Beauty Face Teen O525547819 Call Girls Dubai Young
Dubai Call Girls Beauty Face Teen O525547819 Call Girls Dubai YoungDubai Call Girls Beauty Face Teen O525547819 Call Girls Dubai Young
Dubai Call Girls Beauty Face Teen O525547819 Call Girls Dubai Young
 
❤Jammu Kashmir Call Girls 8617697112 Personal Whatsapp Number 💦✅.
❤Jammu Kashmir Call Girls 8617697112 Personal Whatsapp Number 💦✅.❤Jammu Kashmir Call Girls 8617697112 Personal Whatsapp Number 💦✅.
❤Jammu Kashmir Call Girls 8617697112 Personal Whatsapp Number 💦✅.
 

Slideshare breaking inter layer co-adaptation

  • 1. Masayuki Tanaka Breaking Inter-Layer Co-Adaptation by Classifier Anonymization Ikuro Sato†, Kohta Ishikawa†, Guoqing Liu†, Masayuki Tanaka‡ (ICML2019) † ‡
  • 2. Meta reviewer’s comment …This paper seems to me like a perfect example of a “High Risk High Reward” paper, … Acceptance ratio of ICML2019: 773/3424 = 22.6% We have taken that as a compliment. It is a research! 1
  • 3. What I’m going to talk 𝑥𝑥 Input 𝐹𝐹𝜙𝜙(𝑥𝑥) 𝐶𝐶𝜃𝜃(𝜉𝜉) 𝜂𝜂 Output 𝜉𝜉 Feature Let’s consider a classification task. Feature extractor Classifier + - Feature space 𝜉𝜉 + + + + + + + -- - - -- - - Feature space 𝜉𝜉 + ++ + + + +-- -- -- - End-to-end DNN << Which is better? Why? How can we obtain good features?2
  • 4. Summary About what? How? Theory? In reality? Breaking co-adaptation between feature extractor and classifier. By classifier anonymization technique. Proved: Features form simple point-like distribution. Point-like property largely confirmed on real datasets. 3
  • 5. What is a co-adaptation? 𝑥𝑥 Input 𝐹𝐹𝜙𝜙(𝑥𝑥) 𝐶𝐶𝜃𝜃(𝜉𝜉) 𝜂𝜂 Output 𝜉𝜉 Feature Let’s consider a classification task. Feature extractor Classifier + - Feature space 𝜉𝜉 Decision boundary + + + + + + + -- - - -- - Co-adaptation: Feature extractor adapts a particular classifier. Classifier adapts a particular feature extractor. Break co-adaptation - Feature space 𝜉𝜉 + ++ + + + +-- -- -- - Classifiers Feature extractor should be trained for many classifiers. End-to-end DNN 4
  • 6. Proposed algorithm: FOCA - Feature space 𝜉𝜉 +++ + + ++ -- ----- (Under several conditions,) we theoretically proved the FOCA can train the feature extractor which projects single point. for given feature extractor FOCA can train feature extractor to make any weak classifier strong. FOCA: Feature-extractor Optimization through Classifier Anonymization 5
  • 7. Message of FOCA Traditional training FOCA training Feature extractor (Junior researcher) Feature extractor (Junior researcher) Weak classifiers (Boss variety???) Strong classifier (Smart boss) Transfer learning (New boss, new domain) FOCA can train feature extractor strong. 6
  • 8. Weak classifier assumption Definition: Weak classifier is slightly better than random guess. 𝜃𝜃𝜙𝜙 ∗ = arg min 𝜃𝜃 E (𝑥𝑥,𝑡𝑡)~𝑝𝑝(𝑥𝑥,𝑡𝑡) 𝐿𝐿 𝐶𝐶𝜃𝜃 𝐹𝐹𝜙𝜙(𝑥𝑥) , 𝑡𝑡 𝜃𝜃𝜙𝜙 𝐵𝐵 = arg min 𝜃𝜃 � 𝑥𝑥,𝑡𝑡 ∈𝐵𝐵 𝐿𝐿 𝐶𝐶𝜃𝜃 𝐹𝐹𝜙𝜙(𝑥𝑥) , 𝑡𝑡 Strong classifier Strong classifier is strong for entire data. Weak classifier assumption We assume that strong classifier for small samples is weak classifier for entire data. B is small samples of entire data. 7
  • 9. Practical FOCA algorithm 𝐹𝐹𝜙𝜙(𝑥𝑥) 𝐶𝐶𝜃𝜃(𝜉𝜉) Weak classifier generatorFeature extractor Classifier model 𝐹𝐹𝐹𝜙𝜙(𝑥𝑥) Previous feature extractor Training data Optimize the classifier for given small samples with previous feature extractor. Update feature extractor for given mini-batch with weak classifier. Sampling 𝐶𝐶𝜃𝜃(𝜉𝜉) Weak classifier Update Mini-batch 8
  • 10. Experimental validation Two-step training: Train the feature extractor. Then, train the classifier with the fixed given feature extractor. - Feature space 𝜉𝜉 + + + + + + + -- - - -- - Co-adaptation Point-like - Feature space 𝜉𝜉 +++ + + ++ -- ----- Many samples are required to train the classifier. A few samples are good enough to train the classifier. 9
  • 12. Poster as a summary 11
  • 13. Links Official proceedings of ICML2019 http://proceedings.mlr.press/v97/ arxiv: Breaking Inter-Layer Co-Adaptation by Classifier Anonymization https://arxiv.org/abs/1906.01150 Twitter: Masayuki Tanaka https://twitter.com/likesilkto Twitter: Ikuro Sato https://twitter.com/ikuro_s 12