SlideShare a Scribd company logo
Visual Causal
Feature
Learning
Kojin Oshiba
Paper
Overview
Theory of
Visual Causal
Learning
Causal Feature
Learning
Algorithm
Experiment
Discussion
Visual Causal Feature Learning
Chalupka et al. 2015
Kojin Oshiba
Department of Computer Science & Department of Statistics
Harvard University
October 12, 2018
Visual Causal
Feature
Learning
Kojin Oshiba
Paper
Overview
Theory of
Visual Causal
Learning
Causal Feature
Learning
Algorithm
Experiment
Discussion
Table of Contents
1 Paper Overview
2 Theory of Visual Causal Learning
3 Causal Feature Learning Algorithm
4 Experiment
5 Discussion
Visual Causal
Feature
Learning
Kojin Oshiba
Paper
Overview
Theory of
Visual Causal
Learning
Causal Feature
Learning
Algorithm
Experiment
Discussion
Table of Contents
1 Paper Overview
2 Theory of Visual Causal Learning
3 Causal Feature Learning Algorithm
4 Experiment
5 Discussion
Visual Causal
Feature
Learning
Kojin Oshiba
Paper
Overview
Theory of
Visual Causal
Learning
Causal Feature
Learning
Algorithm
Experiment
Discussion
Overview
Goal: Understanding the visual cause of humans.
• A framework for causal learning from macro variables (e.g. groups of pixels).
• Observational data + minimal experiment = visual cause
• Applicable to any aggregate of micro variables, e.g. auditory, olfactory data.
Visual Causal
Feature
Learning
Kojin Oshiba
Paper
Overview
Theory of
Visual Causal
Learning
Causal Feature
Learning
Algorithm
Experiment
Discussion
Overview
Goal: Understanding the visual cause of humans.
• A framework for causal learning from macro variables (e.g. groups of pixels).
• Observational data + minimal experiment = visual cause
• Applicable to any aggregate of micro variables, e.g. auditory, olfactory data.
Technically,
• Define a macro-variable C, which contains all the causal information available
in an image I ∈ I about a given behavior T.
• I is generated from unobserved discrete variables H.
Visual Causal
Feature
Learning
Kojin Oshiba
Paper
Overview
Theory of
Visual Causal
Learning
Causal Feature
Learning
Algorithm
Experiment
Discussion
Table of Contents
1 Paper Overview
2 Theory of Visual Causal Learning
3 Causal Feature Learning Algorithm
4 Experiment
5 Discussion
Visual Causal
Feature
Learning
Kojin Oshiba
Paper
Overview
Theory of
Visual Causal
Learning
Causal Feature
Learning
Algorithm
Experiment
Discussion
Observational Partition
Definition: Observational Partition
Πo(T, I) ⊂ I is a partition based on the equivalence relation
i ∼ j ⇔ P(T|I = i) = P(T|I = j) ∀i, j ∈ Πo(T, I)
Knowing the observational partition of an image allows us to predict the value of T.
Visual Causal
Feature
Learning
Kojin Oshiba
Paper
Overview
Theory of
Visual Causal
Learning
Causal Feature
Learning
Algorithm
Experiment
Discussion
Causal Partition
Definition: Causal Partition
Πc(T, I) ⊂ I, is a partition based on the equivalence relation
i ∼ j ⇔ P(T|man(I = i)) = P(T|man(I = j)).
Manipulation changes the image, but not T or H.
Visual Causal
Feature
Learning
Kojin Oshiba
Paper
Overview
Theory of
Visual Causal
Learning
Causal Feature
Learning
Algorithm
Experiment
Discussion
Causal Partition
Definition: Causal Partition
Πc(T, I) ⊂ I, is a partition based on the equivalence relation
i ∼ j ⇔ P(T|man(I = i)) = P(T|man(I = j)).
Manipulation changes the image, but not T or H.
Definition: Visual Cause
C(i) = P(T|man(I = i)).
Has one-to-one correspondence with Πc(T, I).
Visual Causal
Feature
Learning
Kojin Oshiba
Paper
Overview
Theory of
Visual Causal
Learning
Causal Feature
Learning
Algorithm
Experiment
Discussion
Causal Coarsening Theorem
Theorem: Causal Coarsening
Πc is a coarsening of Πo, i.e. causal labels do not change within each
observational class.
Visual Causal
Feature
Learning
Kojin Oshiba
Paper
Overview
Theory of
Visual Causal
Learning
Causal Feature
Learning
Algorithm
Experiment
Discussion
Causal Coarsening Theorem
Theorem: Causal Coarsening
Πc is a coarsening of Πo, i.e. causal labels do not change within each
observational class.
The visual causes do not contain all the information in the image that predict T.
∃ information, not itself causal, can be informative about non-visual causes of T.
This is called the spurious correlate.
Visual Causal
Feature
Learning
Kojin Oshiba
Paper
Overview
Theory of
Visual Causal
Learning
Causal Feature
Learning
Algorithm
Experiment
Discussion
Causal Coarsening Theorem
Theorem: Causal Coarsening
Πc is a coarsening of Πo, i.e. causal labels do not change within each
observational class.
The visual causes do not contain all the information in the image that predict T.
∃ information, not itself causal, can be informative about non-visual causes of T.
This is called the spurious correlate.
Definition: Spurrious Correlate
S is a discrete random variable whose value differentiates between Πo(T, I)
contained in Πc(T, I).
Visual Causal
Feature
Learning
Kojin Oshiba
Paper
Overview
Theory of
Visual Causal
Learning
Causal Feature
Learning
Algorithm
Experiment
Discussion
Causal Coarsening Theorem
Theorem: Causal Coarsening
Πc is a coarsening of Πo, i.e. causal labels do not change within each
observational class.
The visual causes do not contain all the information in the image that predict T.
∃ information, not itself causal, can be informative about non-visual causes of T.
This is called the spurious correlate.
Definition: Spurrious Correlate
S is a discrete random variable whose value differentiates between Πo(T, I)
contained in Πc(T, I).
Theorem: Complete Macro-variable Description
C and S together contain all and only the visual information in I relevant to T,
but only C contains the causal information
Visual Causal
Feature
Learning
Kojin Oshiba
Paper
Overview
Theory of
Visual Causal
Learning
Causal Feature
Learning
Algorithm
Experiment
Discussion
Causal Intervention on Macro-variables
Definition: Visual Manipulation man(I = i)
An operation that changes (the pixels of) the image to image i ∈ I, while not
affecting any other variables (such as H or T).
Visual Causal
Feature
Learning
Kojin Oshiba
Paper
Overview
Theory of
Visual Causal
Learning
Causal Feature
Learning
Algorithm
Experiment
Discussion
Causal Intervention on Macro-variables
Definition: Visual Manipulation man(I = i)
An operation that changes (the pixels of) the image to image i ∈ I, while not
affecting any other variables (such as H or T).
Definition: Causal Intervention on Macro-variables do(C = c )
man(I = i ) such that C(i ) = c and S(i ) = s.
This is not always possible.
Visual Causal
Feature
Learning
Kojin Oshiba
Paper
Overview
Theory of
Visual Causal
Learning
Causal Feature
Learning
Algorithm
Experiment
Discussion
Causal Intervention on Macro-variables
Definition: Visual Manipulation man(I = i)
An operation that changes (the pixels of) the image to image i ∈ I, while not
affecting any other variables (such as H or T).
Definition: Causal Intervention on Macro-variables do(C = c )
man(I = i ) such that C(i ) = c and S(i ) = s.
This is not always possible.
Phew! We have the theory down. Now the question is: how can we learn C?
Visual Causal
Feature
Learning
Kojin Oshiba
Paper
Overview
Theory of
Visual Causal
Learning
Causal Feature
Learning
Algorithm
Experiment
Discussion
Table of Contents
1 Paper Overview
2 Theory of Visual Causal Learning
3 Causal Feature Learning Algorithm
4 Experiment
5 Discussion
Visual Causal
Feature
Learning
Kojin Oshiba
Paper
Overview
Theory of
Visual Causal
Learning
Causal Feature
Learning
Algorithm
Experiment
Discussion
Causal Effect Prediction
Steps:
0 Observational class for each data is
necessary (do we estimate?).
1 Picks a representative member of each
observational class.
2 Need to design an experiment here.
3 Coarsen observational partitions to
construct causal paritions.
4 Training a neural net on this imputed data
gives the causal effect.
5 NN is our C.
Visual Causal
Feature
Learning
Kojin Oshiba
Paper
Overview
Theory of
Visual Causal
Learning
Causal Feature
Learning
Algorithm
Experiment
Discussion
Causal Feature Manipulation
Steps:
0 Oracle A can be obtained from Algo 1.
1 Train manipulation function every iter.
2 Train a "causal neural net".
3 Choose images to be manipulated.
4 Choose target causal partition.
5 For each data,
6 Since C is hardly invertible, this
approximates argminˆi∈C−1(k) d(i, ˆi)
7
8 Augment the causal data every iteration.
9 I think this algo doesn’t really learn MC but
this algo itself is MC ...
Visual Causal
Feature
Learning
Kojin Oshiba
Paper
Overview
Theory of
Visual Causal
Learning
Causal Feature
Learning
Algorithm
Experiment
Discussion
Table of Contents
1 Paper Overview
2 Theory of Visual Causal Learning
3 Causal Feature Learning Algorithm
4 Experiment
5 Discussion
Visual Causal
Feature
Learning
Kojin Oshiba
Paper
Overview
Theory of
Visual Causal
Learning
Causal Feature
Learning
Algorithm
Experiment
Discussion
Causal Intervention on Macro-variables
• T = 1 if a human answers affirmatively to
the question "does this image contain the
digit ’x’?", where x is the actual digit on
the image.
• Assume for simplicity that
P(T = 1|man(I)) = 0 or
P(T = 1|man(I)) = 1.
• Here, they already know the causal data:
the labels are assigned by default. There
were no experiments training Algo 1, so
unclear if it works in practice.
Visual Causal
Feature
Learning
Kojin Oshiba
Paper
Overview
Theory of
Visual Causal
Learning
Causal Feature
Learning
Algorithm
Experiment
Discussion
Table of Contents
1 Paper Overview
2 Theory of Visual Causal Learning
3 Causal Feature Learning Algorithm
4 Experiment
5 Discussion
Visual Causal
Feature
Learning
Kojin Oshiba
Paper
Overview
Theory of
Visual Causal
Learning
Causal Feature
Learning
Algorithm
Experiment
Discussion
Other Types of "Causality" on Visual Data
• Understanding the causal structure within an image: Discovering Causal
Signals in Images (Lopez-Paz et al. 2017), CausalGAN (Kocaoglu et al. 2017)
Visual Causal
Feature
Learning
Kojin Oshiba
Paper
Overview
Theory of
Visual Causal
Learning
Causal Feature
Learning
Algorithm
Experiment
Discussion
Fin.

More Related Content

What's hot

Introduction.doc
Introduction.docIntroduction.doc
Introduction.doc
butest
 
Doubt intuitionistic fuzzy deals in bckbci algebras
Doubt intuitionistic fuzzy deals in bckbci algebrasDoubt intuitionistic fuzzy deals in bckbci algebras
Doubt intuitionistic fuzzy deals in bckbci algebras
ijfls
 
Bayesian inference for mixed-effects models driven by SDEs and other stochast...
Bayesian inference for mixed-effects models driven by SDEs and other stochast...Bayesian inference for mixed-effects models driven by SDEs and other stochast...
Bayesian inference for mixed-effects models driven by SDEs and other stochast...
Umberto Picchini
 
A Method for Solving Balanced Intuitionistic Fuzzy Assignment Problem
A  Method  for  Solving  Balanced  Intuitionistic  Fuzzy  Assignment  Problem A  Method  for  Solving  Balanced  Intuitionistic  Fuzzy  Assignment  Problem
A Method for Solving Balanced Intuitionistic Fuzzy Assignment Problem
Navodaya Institute of Technology
 
Ravens intelligence test
Ravens intelligence testRavens intelligence test
Ev4301897903
Ev4301897903Ev4301897903
Ev4301897903
IJERA Editor
 
Intuitionistic Fuzzy W- Closed Sets and Intuitionistic Fuzzy W -Continuity
Intuitionistic Fuzzy W- Closed Sets and Intuitionistic Fuzzy W -ContinuityIntuitionistic Fuzzy W- Closed Sets and Intuitionistic Fuzzy W -Continuity
Intuitionistic Fuzzy W- Closed Sets and Intuitionistic Fuzzy W -Continuity
Waqas Tariq
 
Large-Scale Nonparametric Estimation of Vehicle Travel Time Distributions
Large-Scale Nonparametric Estimation of Vehicle Travel Time DistributionsLarge-Scale Nonparametric Estimation of Vehicle Travel Time Distributions
Large-Scale Nonparametric Estimation of Vehicle Travel Time Distributions
Rikiya Takahashi
 
Automatic Image Annotation (AIA)
Automatic Image Annotation (AIA)Automatic Image Annotation (AIA)
Automatic Image Annotation (AIA)
Farzaneh Rezaei
 
Recognizing Human-Object Interactions in Still Images by Modeling the Mutual ...
Recognizing Human-Object Interactions inStill Images by Modeling the Mutual ...Recognizing Human-Object Interactions inStill Images by Modeling the Mutual ...
Recognizing Human-Object Interactions in Still Images by Modeling the Mutual ...
أحلام انصارى
 

What's hot (10)

Introduction.doc
Introduction.docIntroduction.doc
Introduction.doc
 
Doubt intuitionistic fuzzy deals in bckbci algebras
Doubt intuitionistic fuzzy deals in bckbci algebrasDoubt intuitionistic fuzzy deals in bckbci algebras
Doubt intuitionistic fuzzy deals in bckbci algebras
 
Bayesian inference for mixed-effects models driven by SDEs and other stochast...
Bayesian inference for mixed-effects models driven by SDEs and other stochast...Bayesian inference for mixed-effects models driven by SDEs and other stochast...
Bayesian inference for mixed-effects models driven by SDEs and other stochast...
 
A Method for Solving Balanced Intuitionistic Fuzzy Assignment Problem
A  Method  for  Solving  Balanced  Intuitionistic  Fuzzy  Assignment  Problem A  Method  for  Solving  Balanced  Intuitionistic  Fuzzy  Assignment  Problem
A Method for Solving Balanced Intuitionistic Fuzzy Assignment Problem
 
Ravens intelligence test
Ravens intelligence testRavens intelligence test
Ravens intelligence test
 
Ev4301897903
Ev4301897903Ev4301897903
Ev4301897903
 
Intuitionistic Fuzzy W- Closed Sets and Intuitionistic Fuzzy W -Continuity
Intuitionistic Fuzzy W- Closed Sets and Intuitionistic Fuzzy W -ContinuityIntuitionistic Fuzzy W- Closed Sets and Intuitionistic Fuzzy W -Continuity
Intuitionistic Fuzzy W- Closed Sets and Intuitionistic Fuzzy W -Continuity
 
Large-Scale Nonparametric Estimation of Vehicle Travel Time Distributions
Large-Scale Nonparametric Estimation of Vehicle Travel Time DistributionsLarge-Scale Nonparametric Estimation of Vehicle Travel Time Distributions
Large-Scale Nonparametric Estimation of Vehicle Travel Time Distributions
 
Automatic Image Annotation (AIA)
Automatic Image Annotation (AIA)Automatic Image Annotation (AIA)
Automatic Image Annotation (AIA)
 
Recognizing Human-Object Interactions in Still Images by Modeling the Mutual ...
Recognizing Human-Object Interactions inStill Images by Modeling the Mutual ...Recognizing Human-Object Interactions inStill Images by Modeling the Mutual ...
Recognizing Human-Object Interactions in Still Images by Modeling the Mutual ...
 

Similar to Visutl Causal Feature Learning

AET vs. AED: Unsupervised Representation Learning by Auto-Encoding Transforma...
AET vs. AED: Unsupervised Representation Learning by Auto-Encoding Transforma...AET vs. AED: Unsupervised Representation Learning by Auto-Encoding Transforma...
AET vs. AED: Unsupervised Representation Learning by Auto-Encoding Transforma...
Tomoyuki Suzuki
 
Presentation
PresentationPresentation
Presentation
Himani Himmi
 
Seminar CCC
Seminar CCCSeminar CCC
Seminar CCC
Antonio Lieto
 
Commonsense reasoning as a key feature for dynamic knowledge invention and co...
Commonsense reasoning as a key feature for dynamic knowledge invention and co...Commonsense reasoning as a key feature for dynamic knowledge invention and co...
Commonsense reasoning as a key feature for dynamic knowledge invention and co...
Antonio Lieto
 
Interpretability of machine learning
Interpretability of machine learningInterpretability of machine learning
Interpretability of machine learning
Daiki Tanaka
 
Mit6870 orsu lecture2
Mit6870 orsu lecture2Mit6870 orsu lecture2
Mit6870 orsu lecture2
zukun
 
A likelihood-free version of the stochastic approximation EM algorithm (SAEM)...
A likelihood-free version of the stochastic approximation EM algorithm (SAEM)...A likelihood-free version of the stochastic approximation EM algorithm (SAEM)...
A likelihood-free version of the stochastic approximation EM algorithm (SAEM)...
Umberto Picchini
 
Visual Analytics in Omics: why, what, how?
Visual Analytics in Omics: why, what, how?Visual Analytics in Omics: why, what, how?
Visual Analytics in Omics: why, what, how?
Jan Aerts
 
[DL輪読会]Generative Models of Visually Grounded Imagination
[DL輪読会]Generative Models of Visually Grounded Imagination[DL輪読会]Generative Models of Visually Grounded Imagination
[DL輪読会]Generative Models of Visually Grounded Imagination
Deep Learning JP
 
Knowledge Capturing via Conceptual Reframing: A Goal-oriented Framework for K...
Knowledge Capturing via Conceptual Reframing: A Goal-oriented Framework for K...Knowledge Capturing via Conceptual Reframing: A Goal-oriented Framework for K...
Knowledge Capturing via Conceptual Reframing: A Goal-oriented Framework for K...
Antonio Lieto
 
A measure to evaluate latent variable model fit by sensitivity analysis
A measure to evaluate latent variable model fit by sensitivity analysisA measure to evaluate latent variable model fit by sensitivity analysis
A measure to evaluate latent variable model fit by sensitivity analysis
Daniel Oberski
 
nncollovcapaldo2013-131220052427-phpapp01.pdf
nncollovcapaldo2013-131220052427-phpapp01.pdfnncollovcapaldo2013-131220052427-phpapp01.pdf
nncollovcapaldo2013-131220052427-phpapp01.pdf
GayathriRHICETCSESTA
 
nncollovcapaldo2013-131220052427-phpapp01.pdf
nncollovcapaldo2013-131220052427-phpapp01.pdfnncollovcapaldo2013-131220052427-phpapp01.pdf
nncollovcapaldo2013-131220052427-phpapp01.pdf
GayathriRHICETCSESTA
 
Learning
LearningLearning
Learning
butest
 
Intro to Approximate Bayesian Computation (ABC)
Intro to Approximate Bayesian Computation (ABC)Intro to Approximate Bayesian Computation (ABC)
Intro to Approximate Bayesian Computation (ABC)
Umberto Picchini
 
scene description
scene descriptionscene description
scene description
khushi2551
 
Epistemologia computacional: intrudução
Epistemologia computacional: intruduçãoEpistemologia computacional: intrudução
Epistemologia computacional: intrudução
Danilo Fraga Dantas
 
SIP-REVIEW-3.pptx
SIP-REVIEW-3.pptxSIP-REVIEW-3.pptx
SIP-REVIEW-3.pptx
rajhumdabad
 
TPCMFinalACone
TPCMFinalAConeTPCMFinalACone
TPCMFinalACone
Adam Cone
 
Lecture17 xing fei-fei
Lecture17 xing fei-feiLecture17 xing fei-fei
Lecture17 xing fei-fei
Tianlu Wang
 

Similar to Visutl Causal Feature Learning (20)

AET vs. AED: Unsupervised Representation Learning by Auto-Encoding Transforma...
AET vs. AED: Unsupervised Representation Learning by Auto-Encoding Transforma...AET vs. AED: Unsupervised Representation Learning by Auto-Encoding Transforma...
AET vs. AED: Unsupervised Representation Learning by Auto-Encoding Transforma...
 
Presentation
PresentationPresentation
Presentation
 
Seminar CCC
Seminar CCCSeminar CCC
Seminar CCC
 
Commonsense reasoning as a key feature for dynamic knowledge invention and co...
Commonsense reasoning as a key feature for dynamic knowledge invention and co...Commonsense reasoning as a key feature for dynamic knowledge invention and co...
Commonsense reasoning as a key feature for dynamic knowledge invention and co...
 
Interpretability of machine learning
Interpretability of machine learningInterpretability of machine learning
Interpretability of machine learning
 
Mit6870 orsu lecture2
Mit6870 orsu lecture2Mit6870 orsu lecture2
Mit6870 orsu lecture2
 
A likelihood-free version of the stochastic approximation EM algorithm (SAEM)...
A likelihood-free version of the stochastic approximation EM algorithm (SAEM)...A likelihood-free version of the stochastic approximation EM algorithm (SAEM)...
A likelihood-free version of the stochastic approximation EM algorithm (SAEM)...
 
Visual Analytics in Omics: why, what, how?
Visual Analytics in Omics: why, what, how?Visual Analytics in Omics: why, what, how?
Visual Analytics in Omics: why, what, how?
 
[DL輪読会]Generative Models of Visually Grounded Imagination
[DL輪読会]Generative Models of Visually Grounded Imagination[DL輪読会]Generative Models of Visually Grounded Imagination
[DL輪読会]Generative Models of Visually Grounded Imagination
 
Knowledge Capturing via Conceptual Reframing: A Goal-oriented Framework for K...
Knowledge Capturing via Conceptual Reframing: A Goal-oriented Framework for K...Knowledge Capturing via Conceptual Reframing: A Goal-oriented Framework for K...
Knowledge Capturing via Conceptual Reframing: A Goal-oriented Framework for K...
 
A measure to evaluate latent variable model fit by sensitivity analysis
A measure to evaluate latent variable model fit by sensitivity analysisA measure to evaluate latent variable model fit by sensitivity analysis
A measure to evaluate latent variable model fit by sensitivity analysis
 
nncollovcapaldo2013-131220052427-phpapp01.pdf
nncollovcapaldo2013-131220052427-phpapp01.pdfnncollovcapaldo2013-131220052427-phpapp01.pdf
nncollovcapaldo2013-131220052427-phpapp01.pdf
 
nncollovcapaldo2013-131220052427-phpapp01.pdf
nncollovcapaldo2013-131220052427-phpapp01.pdfnncollovcapaldo2013-131220052427-phpapp01.pdf
nncollovcapaldo2013-131220052427-phpapp01.pdf
 
Learning
LearningLearning
Learning
 
Intro to Approximate Bayesian Computation (ABC)
Intro to Approximate Bayesian Computation (ABC)Intro to Approximate Bayesian Computation (ABC)
Intro to Approximate Bayesian Computation (ABC)
 
scene description
scene descriptionscene description
scene description
 
Epistemologia computacional: intrudução
Epistemologia computacional: intruduçãoEpistemologia computacional: intrudução
Epistemologia computacional: intrudução
 
SIP-REVIEW-3.pptx
SIP-REVIEW-3.pptxSIP-REVIEW-3.pptx
SIP-REVIEW-3.pptx
 
TPCMFinalACone
TPCMFinalAConeTPCMFinalACone
TPCMFinalACone
 
Lecture17 xing fei-fei
Lecture17 xing fei-feiLecture17 xing fei-fei
Lecture17 xing fei-fei
 

Recently uploaded

Intelligence supported media monitoring in veterinary medicine
Intelligence supported media monitoring in veterinary medicineIntelligence supported media monitoring in veterinary medicine
Intelligence supported media monitoring in veterinary medicine
AndrzejJarynowski
 
Beyond the Basics of A/B Tests: Highly Innovative Experimentation Tactics You...
Beyond the Basics of A/B Tests: Highly Innovative Experimentation Tactics You...Beyond the Basics of A/B Tests: Highly Innovative Experimentation Tactics You...
Beyond the Basics of A/B Tests: Highly Innovative Experimentation Tactics You...
Aggregage
 
一比一原版(GWU,GW文凭证书)乔治·华盛顿大学毕业证如何办理
一比一原版(GWU,GW文凭证书)乔治·华盛顿大学毕业证如何办理一比一原版(GWU,GW文凭证书)乔治·华盛顿大学毕业证如何办理
一比一原版(GWU,GW文凭证书)乔治·华盛顿大学毕业证如何办理
bopyb
 
Population Growth in Bataan: The effects of population growth around rural pl...
Population Growth in Bataan: The effects of population growth around rural pl...Population Growth in Bataan: The effects of population growth around rural pl...
Population Growth in Bataan: The effects of population growth around rural pl...
Bill641377
 
06-04-2024 - NYC Tech Week - Discussion on Vector Databases, Unstructured Dat...
06-04-2024 - NYC Tech Week - Discussion on Vector Databases, Unstructured Dat...06-04-2024 - NYC Tech Week - Discussion on Vector Databases, Unstructured Dat...
06-04-2024 - NYC Tech Week - Discussion on Vector Databases, Unstructured Dat...
Timothy Spann
 
Everything you wanted to know about LIHTC
Everything you wanted to know about LIHTCEverything you wanted to know about LIHTC
Everything you wanted to know about LIHTC
Roger Valdez
 
Learn SQL from basic queries to Advance queries
Learn SQL from basic queries to Advance queriesLearn SQL from basic queries to Advance queries
Learn SQL from basic queries to Advance queries
manishkhaire30
 
一比一原版(Glasgow毕业证书)格拉斯哥大学毕业证如何办理
一比一原版(Glasgow毕业证书)格拉斯哥大学毕业证如何办理一比一原版(Glasgow毕业证书)格拉斯哥大学毕业证如何办理
一比一原版(Glasgow毕业证书)格拉斯哥大学毕业证如何办理
g4dpvqap0
 
Experts live - Improving user adoption with AI
Experts live - Improving user adoption with AIExperts live - Improving user adoption with AI
Experts live - Improving user adoption with AI
jitskeb
 
ViewShift: Hassle-free Dynamic Policy Enforcement for Every Data Lake
ViewShift: Hassle-free Dynamic Policy Enforcement for Every Data LakeViewShift: Hassle-free Dynamic Policy Enforcement for Every Data Lake
ViewShift: Hassle-free Dynamic Policy Enforcement for Every Data Lake
Walaa Eldin Moustafa
 
一比一原版(UCSF文凭证书)旧金山分校毕业证如何办理
一比一原版(UCSF文凭证书)旧金山分校毕业证如何办理一比一原版(UCSF文凭证书)旧金山分校毕业证如何办理
一比一原版(UCSF文凭证书)旧金山分校毕业证如何办理
nuttdpt
 
STATATHON: Unleashing the Power of Statistics in a 48-Hour Knowledge Extravag...
STATATHON: Unleashing the Power of Statistics in a 48-Hour Knowledge Extravag...STATATHON: Unleashing the Power of Statistics in a 48-Hour Knowledge Extravag...
STATATHON: Unleashing the Power of Statistics in a 48-Hour Knowledge Extravag...
sameer shah
 
一比一原版(爱大毕业证书)爱丁堡大学毕业证如何办理
一比一原版(爱大毕业证书)爱丁堡大学毕业证如何办理一比一原版(爱大毕业证书)爱丁堡大学毕业证如何办理
一比一原版(爱大毕业证书)爱丁堡大学毕业证如何办理
g4dpvqap0
 
State of Artificial intelligence Report 2023
State of Artificial intelligence Report 2023State of Artificial intelligence Report 2023
State of Artificial intelligence Report 2023
kuntobimo2016
 
The Ipsos - AI - Monitor 2024 Report.pdf
The  Ipsos - AI - Monitor 2024 Report.pdfThe  Ipsos - AI - Monitor 2024 Report.pdf
The Ipsos - AI - Monitor 2024 Report.pdf
Social Samosa
 
My burning issue is homelessness K.C.M.O.
My burning issue is homelessness K.C.M.O.My burning issue is homelessness K.C.M.O.
My burning issue is homelessness K.C.M.O.
rwarrenll
 
一比一原版(UO毕业证)渥太华大学毕业证如何办理
一比一原版(UO毕业证)渥太华大学毕业证如何办理一比一原版(UO毕业证)渥太华大学毕业证如何办理
一比一原版(UO毕业证)渥太华大学毕业证如何办理
aqzctr7x
 
Udemy_2024_Global_Learning_Skills_Trends_Report (1).pdf
Udemy_2024_Global_Learning_Skills_Trends_Report (1).pdfUdemy_2024_Global_Learning_Skills_Trends_Report (1).pdf
Udemy_2024_Global_Learning_Skills_Trends_Report (1).pdf
Fernanda Palhano
 
在线办理(英国UCA毕业证书)创意艺术大学毕业证在读证明一模一样
在线办理(英国UCA毕业证书)创意艺术大学毕业证在读证明一模一样在线办理(英国UCA毕业证书)创意艺术大学毕业证在读证明一模一样
在线办理(英国UCA毕业证书)创意艺术大学毕业证在读证明一模一样
v7oacc3l
 
一比一原版(Coventry毕业证书)考文垂大学毕业证如何办理
一比一原版(Coventry毕业证书)考文垂大学毕业证如何办理一比一原版(Coventry毕业证书)考文垂大学毕业证如何办理
一比一原版(Coventry毕业证书)考文垂大学毕业证如何办理
74nqk8xf
 

Recently uploaded (20)

Intelligence supported media monitoring in veterinary medicine
Intelligence supported media monitoring in veterinary medicineIntelligence supported media monitoring in veterinary medicine
Intelligence supported media monitoring in veterinary medicine
 
Beyond the Basics of A/B Tests: Highly Innovative Experimentation Tactics You...
Beyond the Basics of A/B Tests: Highly Innovative Experimentation Tactics You...Beyond the Basics of A/B Tests: Highly Innovative Experimentation Tactics You...
Beyond the Basics of A/B Tests: Highly Innovative Experimentation Tactics You...
 
一比一原版(GWU,GW文凭证书)乔治·华盛顿大学毕业证如何办理
一比一原版(GWU,GW文凭证书)乔治·华盛顿大学毕业证如何办理一比一原版(GWU,GW文凭证书)乔治·华盛顿大学毕业证如何办理
一比一原版(GWU,GW文凭证书)乔治·华盛顿大学毕业证如何办理
 
Population Growth in Bataan: The effects of population growth around rural pl...
Population Growth in Bataan: The effects of population growth around rural pl...Population Growth in Bataan: The effects of population growth around rural pl...
Population Growth in Bataan: The effects of population growth around rural pl...
 
06-04-2024 - NYC Tech Week - Discussion on Vector Databases, Unstructured Dat...
06-04-2024 - NYC Tech Week - Discussion on Vector Databases, Unstructured Dat...06-04-2024 - NYC Tech Week - Discussion on Vector Databases, Unstructured Dat...
06-04-2024 - NYC Tech Week - Discussion on Vector Databases, Unstructured Dat...
 
Everything you wanted to know about LIHTC
Everything you wanted to know about LIHTCEverything you wanted to know about LIHTC
Everything you wanted to know about LIHTC
 
Learn SQL from basic queries to Advance queries
Learn SQL from basic queries to Advance queriesLearn SQL from basic queries to Advance queries
Learn SQL from basic queries to Advance queries
 
一比一原版(Glasgow毕业证书)格拉斯哥大学毕业证如何办理
一比一原版(Glasgow毕业证书)格拉斯哥大学毕业证如何办理一比一原版(Glasgow毕业证书)格拉斯哥大学毕业证如何办理
一比一原版(Glasgow毕业证书)格拉斯哥大学毕业证如何办理
 
Experts live - Improving user adoption with AI
Experts live - Improving user adoption with AIExperts live - Improving user adoption with AI
Experts live - Improving user adoption with AI
 
ViewShift: Hassle-free Dynamic Policy Enforcement for Every Data Lake
ViewShift: Hassle-free Dynamic Policy Enforcement for Every Data LakeViewShift: Hassle-free Dynamic Policy Enforcement for Every Data Lake
ViewShift: Hassle-free Dynamic Policy Enforcement for Every Data Lake
 
一比一原版(UCSF文凭证书)旧金山分校毕业证如何办理
一比一原版(UCSF文凭证书)旧金山分校毕业证如何办理一比一原版(UCSF文凭证书)旧金山分校毕业证如何办理
一比一原版(UCSF文凭证书)旧金山分校毕业证如何办理
 
STATATHON: Unleashing the Power of Statistics in a 48-Hour Knowledge Extravag...
STATATHON: Unleashing the Power of Statistics in a 48-Hour Knowledge Extravag...STATATHON: Unleashing the Power of Statistics in a 48-Hour Knowledge Extravag...
STATATHON: Unleashing the Power of Statistics in a 48-Hour Knowledge Extravag...
 
一比一原版(爱大毕业证书)爱丁堡大学毕业证如何办理
一比一原版(爱大毕业证书)爱丁堡大学毕业证如何办理一比一原版(爱大毕业证书)爱丁堡大学毕业证如何办理
一比一原版(爱大毕业证书)爱丁堡大学毕业证如何办理
 
State of Artificial intelligence Report 2023
State of Artificial intelligence Report 2023State of Artificial intelligence Report 2023
State of Artificial intelligence Report 2023
 
The Ipsos - AI - Monitor 2024 Report.pdf
The  Ipsos - AI - Monitor 2024 Report.pdfThe  Ipsos - AI - Monitor 2024 Report.pdf
The Ipsos - AI - Monitor 2024 Report.pdf
 
My burning issue is homelessness K.C.M.O.
My burning issue is homelessness K.C.M.O.My burning issue is homelessness K.C.M.O.
My burning issue is homelessness K.C.M.O.
 
一比一原版(UO毕业证)渥太华大学毕业证如何办理
一比一原版(UO毕业证)渥太华大学毕业证如何办理一比一原版(UO毕业证)渥太华大学毕业证如何办理
一比一原版(UO毕业证)渥太华大学毕业证如何办理
 
Udemy_2024_Global_Learning_Skills_Trends_Report (1).pdf
Udemy_2024_Global_Learning_Skills_Trends_Report (1).pdfUdemy_2024_Global_Learning_Skills_Trends_Report (1).pdf
Udemy_2024_Global_Learning_Skills_Trends_Report (1).pdf
 
在线办理(英国UCA毕业证书)创意艺术大学毕业证在读证明一模一样
在线办理(英国UCA毕业证书)创意艺术大学毕业证在读证明一模一样在线办理(英国UCA毕业证书)创意艺术大学毕业证在读证明一模一样
在线办理(英国UCA毕业证书)创意艺术大学毕业证在读证明一模一样
 
一比一原版(Coventry毕业证书)考文垂大学毕业证如何办理
一比一原版(Coventry毕业证书)考文垂大学毕业证如何办理一比一原版(Coventry毕业证书)考文垂大学毕业证如何办理
一比一原版(Coventry毕业证书)考文垂大学毕业证如何办理
 

Visutl Causal Feature Learning

  • 1. Visual Causal Feature Learning Kojin Oshiba Paper Overview Theory of Visual Causal Learning Causal Feature Learning Algorithm Experiment Discussion Visual Causal Feature Learning Chalupka et al. 2015 Kojin Oshiba Department of Computer Science & Department of Statistics Harvard University October 12, 2018
  • 2. Visual Causal Feature Learning Kojin Oshiba Paper Overview Theory of Visual Causal Learning Causal Feature Learning Algorithm Experiment Discussion Table of Contents 1 Paper Overview 2 Theory of Visual Causal Learning 3 Causal Feature Learning Algorithm 4 Experiment 5 Discussion
  • 3. Visual Causal Feature Learning Kojin Oshiba Paper Overview Theory of Visual Causal Learning Causal Feature Learning Algorithm Experiment Discussion Table of Contents 1 Paper Overview 2 Theory of Visual Causal Learning 3 Causal Feature Learning Algorithm 4 Experiment 5 Discussion
  • 4. Visual Causal Feature Learning Kojin Oshiba Paper Overview Theory of Visual Causal Learning Causal Feature Learning Algorithm Experiment Discussion Overview Goal: Understanding the visual cause of humans. • A framework for causal learning from macro variables (e.g. groups of pixels). • Observational data + minimal experiment = visual cause • Applicable to any aggregate of micro variables, e.g. auditory, olfactory data.
  • 5. Visual Causal Feature Learning Kojin Oshiba Paper Overview Theory of Visual Causal Learning Causal Feature Learning Algorithm Experiment Discussion Overview Goal: Understanding the visual cause of humans. • A framework for causal learning from macro variables (e.g. groups of pixels). • Observational data + minimal experiment = visual cause • Applicable to any aggregate of micro variables, e.g. auditory, olfactory data. Technically, • Define a macro-variable C, which contains all the causal information available in an image I ∈ I about a given behavior T. • I is generated from unobserved discrete variables H.
  • 6. Visual Causal Feature Learning Kojin Oshiba Paper Overview Theory of Visual Causal Learning Causal Feature Learning Algorithm Experiment Discussion Table of Contents 1 Paper Overview 2 Theory of Visual Causal Learning 3 Causal Feature Learning Algorithm 4 Experiment 5 Discussion
  • 7. Visual Causal Feature Learning Kojin Oshiba Paper Overview Theory of Visual Causal Learning Causal Feature Learning Algorithm Experiment Discussion Observational Partition Definition: Observational Partition Πo(T, I) ⊂ I is a partition based on the equivalence relation i ∼ j ⇔ P(T|I = i) = P(T|I = j) ∀i, j ∈ Πo(T, I) Knowing the observational partition of an image allows us to predict the value of T.
  • 8. Visual Causal Feature Learning Kojin Oshiba Paper Overview Theory of Visual Causal Learning Causal Feature Learning Algorithm Experiment Discussion Causal Partition Definition: Causal Partition Πc(T, I) ⊂ I, is a partition based on the equivalence relation i ∼ j ⇔ P(T|man(I = i)) = P(T|man(I = j)). Manipulation changes the image, but not T or H.
  • 9. Visual Causal Feature Learning Kojin Oshiba Paper Overview Theory of Visual Causal Learning Causal Feature Learning Algorithm Experiment Discussion Causal Partition Definition: Causal Partition Πc(T, I) ⊂ I, is a partition based on the equivalence relation i ∼ j ⇔ P(T|man(I = i)) = P(T|man(I = j)). Manipulation changes the image, but not T or H. Definition: Visual Cause C(i) = P(T|man(I = i)). Has one-to-one correspondence with Πc(T, I).
  • 10. Visual Causal Feature Learning Kojin Oshiba Paper Overview Theory of Visual Causal Learning Causal Feature Learning Algorithm Experiment Discussion Causal Coarsening Theorem Theorem: Causal Coarsening Πc is a coarsening of Πo, i.e. causal labels do not change within each observational class.
  • 11. Visual Causal Feature Learning Kojin Oshiba Paper Overview Theory of Visual Causal Learning Causal Feature Learning Algorithm Experiment Discussion Causal Coarsening Theorem Theorem: Causal Coarsening Πc is a coarsening of Πo, i.e. causal labels do not change within each observational class. The visual causes do not contain all the information in the image that predict T. ∃ information, not itself causal, can be informative about non-visual causes of T. This is called the spurious correlate.
  • 12. Visual Causal Feature Learning Kojin Oshiba Paper Overview Theory of Visual Causal Learning Causal Feature Learning Algorithm Experiment Discussion Causal Coarsening Theorem Theorem: Causal Coarsening Πc is a coarsening of Πo, i.e. causal labels do not change within each observational class. The visual causes do not contain all the information in the image that predict T. ∃ information, not itself causal, can be informative about non-visual causes of T. This is called the spurious correlate. Definition: Spurrious Correlate S is a discrete random variable whose value differentiates between Πo(T, I) contained in Πc(T, I).
  • 13. Visual Causal Feature Learning Kojin Oshiba Paper Overview Theory of Visual Causal Learning Causal Feature Learning Algorithm Experiment Discussion Causal Coarsening Theorem Theorem: Causal Coarsening Πc is a coarsening of Πo, i.e. causal labels do not change within each observational class. The visual causes do not contain all the information in the image that predict T. ∃ information, not itself causal, can be informative about non-visual causes of T. This is called the spurious correlate. Definition: Spurrious Correlate S is a discrete random variable whose value differentiates between Πo(T, I) contained in Πc(T, I). Theorem: Complete Macro-variable Description C and S together contain all and only the visual information in I relevant to T, but only C contains the causal information
  • 14. Visual Causal Feature Learning Kojin Oshiba Paper Overview Theory of Visual Causal Learning Causal Feature Learning Algorithm Experiment Discussion Causal Intervention on Macro-variables Definition: Visual Manipulation man(I = i) An operation that changes (the pixels of) the image to image i ∈ I, while not affecting any other variables (such as H or T).
  • 15. Visual Causal Feature Learning Kojin Oshiba Paper Overview Theory of Visual Causal Learning Causal Feature Learning Algorithm Experiment Discussion Causal Intervention on Macro-variables Definition: Visual Manipulation man(I = i) An operation that changes (the pixels of) the image to image i ∈ I, while not affecting any other variables (such as H or T). Definition: Causal Intervention on Macro-variables do(C = c ) man(I = i ) such that C(i ) = c and S(i ) = s. This is not always possible.
  • 16. Visual Causal Feature Learning Kojin Oshiba Paper Overview Theory of Visual Causal Learning Causal Feature Learning Algorithm Experiment Discussion Causal Intervention on Macro-variables Definition: Visual Manipulation man(I = i) An operation that changes (the pixels of) the image to image i ∈ I, while not affecting any other variables (such as H or T). Definition: Causal Intervention on Macro-variables do(C = c ) man(I = i ) such that C(i ) = c and S(i ) = s. This is not always possible. Phew! We have the theory down. Now the question is: how can we learn C?
  • 17. Visual Causal Feature Learning Kojin Oshiba Paper Overview Theory of Visual Causal Learning Causal Feature Learning Algorithm Experiment Discussion Table of Contents 1 Paper Overview 2 Theory of Visual Causal Learning 3 Causal Feature Learning Algorithm 4 Experiment 5 Discussion
  • 18. Visual Causal Feature Learning Kojin Oshiba Paper Overview Theory of Visual Causal Learning Causal Feature Learning Algorithm Experiment Discussion Causal Effect Prediction Steps: 0 Observational class for each data is necessary (do we estimate?). 1 Picks a representative member of each observational class. 2 Need to design an experiment here. 3 Coarsen observational partitions to construct causal paritions. 4 Training a neural net on this imputed data gives the causal effect. 5 NN is our C.
  • 19. Visual Causal Feature Learning Kojin Oshiba Paper Overview Theory of Visual Causal Learning Causal Feature Learning Algorithm Experiment Discussion Causal Feature Manipulation Steps: 0 Oracle A can be obtained from Algo 1. 1 Train manipulation function every iter. 2 Train a "causal neural net". 3 Choose images to be manipulated. 4 Choose target causal partition. 5 For each data, 6 Since C is hardly invertible, this approximates argminˆi∈C−1(k) d(i, ˆi) 7 8 Augment the causal data every iteration. 9 I think this algo doesn’t really learn MC but this algo itself is MC ...
  • 20. Visual Causal Feature Learning Kojin Oshiba Paper Overview Theory of Visual Causal Learning Causal Feature Learning Algorithm Experiment Discussion Table of Contents 1 Paper Overview 2 Theory of Visual Causal Learning 3 Causal Feature Learning Algorithm 4 Experiment 5 Discussion
  • 21. Visual Causal Feature Learning Kojin Oshiba Paper Overview Theory of Visual Causal Learning Causal Feature Learning Algorithm Experiment Discussion Causal Intervention on Macro-variables • T = 1 if a human answers affirmatively to the question "does this image contain the digit ’x’?", where x is the actual digit on the image. • Assume for simplicity that P(T = 1|man(I)) = 0 or P(T = 1|man(I)) = 1. • Here, they already know the causal data: the labels are assigned by default. There were no experiments training Algo 1, so unclear if it works in practice.
  • 22. Visual Causal Feature Learning Kojin Oshiba Paper Overview Theory of Visual Causal Learning Causal Feature Learning Algorithm Experiment Discussion Table of Contents 1 Paper Overview 2 Theory of Visual Causal Learning 3 Causal Feature Learning Algorithm 4 Experiment 5 Discussion
  • 23. Visual Causal Feature Learning Kojin Oshiba Paper Overview Theory of Visual Causal Learning Causal Feature Learning Algorithm Experiment Discussion Other Types of "Causality" on Visual Data • Understanding the causal structure within an image: Discovering Causal Signals in Images (Lopez-Paz et al. 2017), CausalGAN (Kocaoglu et al. 2017)
  • 24. Visual Causal Feature Learning Kojin Oshiba Paper Overview Theory of Visual Causal Learning Causal Feature Learning Algorithm Experiment Discussion Fin.