SlideShare a Scribd company logo
1 of 32
LEARNING AND THE UNKNOWN:
SURVEYING STEPS TOWARD OPEN WORLD RECOGNITION
Terrance E. Boult
IEEE Fellow
El Pomar Prof. of Innovation and Security
University of Colorado Colorado Springs
I’ll post video of this talk at https://github.com/vastab
Bachelor of InnovationTM
University of Colorado Colorado Springs
WHAT “CLASS” IS THIS OBJECT?
 A few will actually know it
 Most will (quickly) think “I
don’t know” – you know
you don’t know which is
OpenSet
 The curious might try to
look it up and get more
data to “learn” it’s a
Ctenophore aka a comb
jelly – i.e. we do open
world learning
Bachelor of InnovationTM
University of Colorado Colorado Springs
TRAFFICKCAM EXAMPLE (AAAI19 PAPER)
HOTELS-50K: A GLOBAL HOTEL RECOGNITION DATASET:
ABBY STYLIANOU, HONG XUAN, MAYA SHENDE, JONATHAN BRANDT, RICHARD SOUVENIR, ROBERT PLESS
Adversaries will try to defeat a system
Bachelor of InnovationTM
University of Colorado Colorado Springs
SURVEILLANCE EXAMPLE: CAR DETECTION/COUNTING
Nature can produce persistent/long-lived unknown inputs
e.g. Ice on camera housing. (Also bug, bird doodoo…)
Bachelor of InnovationTM
University of Colorado Colorado Springs
Gallery
Probe
Match Score =0.72705
Face Verification (mid JANUS IARPA program)
Bachelor of InnovationTM
University of Colorado Colorado Springs
Match Score= 0.99769
Gallery
Probe
Face Verification (mid-JANUS IARPA program, top performer)
L2-Softmax trained Cosine-distance matching
Bachelor of InnovationTM
University of Colorado Colorado Springs
WHAT WENT WRONG?
Absence of Evidence is not Evidence of Absence
Being far from boundaries & training evidence, A of E,
implies high “probabilities” in classifiers such as
SVM or Softmax
The open set/world is full of “unknowns” that
will absent in training!
Bayesian Reasoning cannot help us as we cannot
normalize without the probability of the unknown inputs
Bachelor of InnovationTM
University of Colorado Colorado Springs
The Range of Openness/Unknowns in problems
Multi-class Classification
Face
Verification
Detection
Open S
et
Recognition
Closed Open
? ? ?
?
? ?
? ?
? ?
Training and
testing samples
come from
known classes
Claimed
identity,
possibility for
impostors
One class,
everything else
in the world is
negative
Multiple known
classes,many
unknown
classes
Multi-class Classification
Face
Verification
Detection
Open S
et
Recognition
Closed Open
? ? ?
?
? ?
? ?
? ?
Training and
testing samples
come from
known classes
Claimed
identity,
possibility for
impostors
One class,
everything else
in the world is
negative
Multiple known
classes,many
unknown
classes
Paper has > 90 citations covering OSR from 14 different application areas
Multi-class Classification
Face
Verification
Detection
Open S
et
Recognition
Closed Open
? ? ?
?
? ?
? ?
? ?
Training and
testing samples
come from
known classes
Claimed
identity,
possibility for
impostors
One class,
everything else
in the world is
negative
Multiple known
classes,many
unknown
classes
Bachelor of InnovationTM
University of Colorado Colorado Springs
Thresholding “Probability” vs Open Set
On optimum recognition error and
reject tradeoff. C. Chow, IEEE
Trans. Info. Theory 16(1):41–46.
1970
Bachelor of InnovationTM
University of Colorado Colorado Springs
LEARNING IN THE FACE OF UNKNOWN UNKNOWNS:
FORMALIZATION OF OPEN-SET RECOGNITION
Open Space Risk Empirical Risk/Error
Scheirer et al. TPAMI ‘13
𝑉 is set of Valid Class training samples;
𝐾 is set of Known unknowns (backgrounds)
Bachelor of InnovationTM
University of Colorado Colorado Springs
OPEN SPACE RISK
“open space” is the space far from known samples. A
simple risk model a constant penalty for labeling that
anything other than unknown in a ratio such as:
Bachelor of InnovationTM
University of Colorado Colorado Springs
Algorithms that solve OSR
RBF SVM
WHAT SOLVES OSR?
Any detector that uses pure linear classifiers, linear SVM,
HAAR cascades, or Softmax-based classifiers, will almost
always have an unbounded open set risk, and hence
does not solve OSR even with thresholding.
GMM
SVDD
If they include a
“bias” or Bayesian
normalization they
probably don’t
solve OSR.
WSVM
PI-SVM
KDE
EVM
NNO
Bachelor of InnovationTM
University of Colorado Colorado Springs
“NOVELTY DETECTION, ANOMALY DETECTION AND
DETECTING “OUT OF DISTRIBUTION SAMPLE”
Long history and many many papers on the first 2, while the latter is a
new term for similar ideas or learning with outliers.
Open-Set Recognition ≅
Anomaly/Novelty detection + Multi-class Recognition
Compute
Novelty or
Anomaly score
Is
Outlier
?
”Closed set”
multi-class
Label
Does NOT directly address “open set recognition” but can be used in sequence to
address OSR. However, rarely is open set recognition part of evaluation evaluation.
Bachelor of InnovationTM
University of Colorado Colorado Springs
 Classic machine learning
presumes all classes
known and classifies all
of feature space.
Thresholding vs Open Set
Bachelor of InnovationTM
University of Colorado Colorado Springs
COMPACT ABATING PROBABILITY
W-SVM ~= OneClassRBF * (EVT scaled) Binary RBF SVM
Bachelor of InnovationTM
University of Colorado Colorado Springs
THERE ARE MANY PROBLEM VARIANTS AND SOLUTIONS
 It can be a part of zero/few shot learning
 Xian, Y.; Lampert, C. H.; Schiele, B.; and Akata, Z. 2018. Zero- shot learning-A
comprehensive evaluation of the good, the bad and the ugly. IEEE TPAMI
2018.
 Or Open set clustering/incremental with no labels
 Active Sampling for Open-Set Classification without Initial Annotation Z-Y. Liu
and S.-J Huang AAA19 (Tech Session 1: Weakly Supervised Learning 1 Thu 2-3:30, Coral
Ballroom 3-5)
 Problems where you have to predict scores on unseen data, e.g.
 “An Open-World Extension to Knowledge Graph Completion Models” H. Shah
et al AAAI 19 (Tech Session 3: Thu 10:25-11:25, Coral 2)
Bachelor of InnovationTM
University of Colorado Colorado Springs
• World with Knowns (K) &
• Known Unknowns (KU)
Unknowns Unknowns (UU)
OSR: Recognizes
known classes or
Detect as Unknown
• NU: Novel
Unknowns
Collect &
Label Data
• LU: Labeled
Unknowns
Incremental
Class Learning
OpenSet
Network
Feature
Training
Class
Label
Unknown
Bendale-Boult CVPR15
Open World Learning
Bachelor of InnovationTM
University of Colorado Colorado Springs
EXTREME VALUE MACHINE
 Uses our “margin distribution theorem” to derive EVT-based
non-linear models that are provably Open-set and also can do
Open World/”Incremental”. Can use classic or deep features
Rudd Et Al. TPAMI 18
Bachelor of InnovationTM
University of Colorado Colorado Springs
TOWARDS OPEN-SET DEEP NETWORKS
Abhijit Bendale*, Terrance Boult
Samsung Research America*
University of Colorado of Colorado Springs
Bachelor of InnovationTM
University of Colorado Colorado Springs
OpenMax for Deep Networks
Distance from MAV
Frequency
FC7
FC8
AlexNet
OpenMax
CAP Model using using EVT
On distances from MAV
Bachelor of InnovationTM
University of Colorado Colorado Springs
OPEN-SET DEEP NETWORKS
Model (MAV) Real Image
Softmax: 0.94, baseball
Fooling Image
Softmax: 1.0, baseball
Open-Set Image
Softmax :0.15, baseball
Openmax : 0.94, baseball Openmax: 0.00, baseball
0.95 Unknown
Openmax: 0.17, baseball
: 0.80, Unknown
Model (MAV)
Real Image
Fooling Image
Open-Set Image
Bachelor of InnovationTM
University of Colorado Colorado Springs
OPENMAX ONLY SOMEWHAT BETTER. WE RESEARCHED WHY.
 Lenet++ 2D feature representation
Bachelor of InnovationTM
University of Colorado Colorado Springs
TRAINING/TESTING
 Known Unknowns
NIST Letters
Known:
MNIST
Bachelor of InnovationTM
University of Colorado Colorado Springs
LENET++ trained on MNIST
MNIST test set (Colors)
LENET++ trained on MNIST
MNIST test set (colors) Black = NIST Letters..
Rather than being far away or “outside” the data,
features for unknown inputs generally overlap known classes
Bachelor of InnovationTM
University of Colorado Colorado Springs
OBSERVATION FROM DEFAULT RESPONSE – LEADING TO OUR APPROACH
Observer that there is difference in entropy and magnitude. While Open-Set
limited response outside the ring of data, most of the unknowns had smaller
magnitude.
The NeurIPS18 approach seeks to emphasize that difference.
Bachelor of InnovationTM
University of Colorado Colorado Springs
GOAL OF OUR APPROACH
Bachelor of InnovationTM
University of Colorado Colorado Springs
LENET++ RESPONSES TO KNOWNS AND UNKNOWNS. Colored dots represent test samples
from the ten MNIST classes, while black dots represent samples from unknown unknowns. The dashed
gray-white lines indicate class borders. The figures in the bottom are histograms of network scores for
known (green) and unknowns (red) with logarithmic vertical axis.
Dhamija et al. 18
Bachelor of InnovationTM
University of Colorado Colorado Springs
Dhamija et al. 18
Bachelor of InnovationTM
University of Colorado Colorado Springs
AISTATS ‘11
Bachelor of InnovationTM
University of Colorado Colorado Springs
Matching Deep Features from
VGG2 set at FAR=10-4 says this
pic of TB matches Barack Obama
So does a commercial system
Adversarial Examples show we do NOT understand how deep
network actually work– “Close in input is not close in features”.
Until we do understand we cannot really do open world deep
networks.
Bachelor of InnovationTM
University of Colorado Colorado Springs
CONCLUSIONS
 We cannot anticipate and train for all “unknown inputs”
 Bayesian reasoning cannot help us if we don’t know
probability of ”unknown” inputs occurring.
 Almost all classical classifiers have unbounded risk and make
highly confident errors. OSR tools address both.
 Traditional deep network map unknown ontop of knowns
 Starting to make progress on deep networks, but it’s a area
with lots of research potential. Adversarial Examples show
there is major issues still “unknown.
Bachelor of InnovationTM
University of Colorado Colorado Springs
“Intelligence comes with hard work and
curiosity for the unknown.”
Roberto Llamas
Do not fear the unknown —
join us in taming it.
https://github.com/vastab
Our code is mostly LIBSVM or BSD-3 ”free” licensed.

More Related Content

Similar to AAAI19-Open.pptx

Bug debug keynote - Present problems and future solutions
Bug debug keynote - Present problems and future solutionsBug debug keynote - Present problems and future solutions
Bug debug keynote - Present problems and future solutions
RIA RUI Society
 
How Four Statistical Rules Forecast Who Wins a Competitive Bid
How Four Statistical Rules Forecast Who Wins a Competitive BidHow Four Statistical Rules Forecast Who Wins a Competitive Bid
How Four Statistical Rules Forecast Who Wins a Competitive Bid
IntelCollab.com
 

Similar to AAAI19-Open.pptx (20)

Bug debug keynote - Present problems and future solutions
Bug debug keynote - Present problems and future solutionsBug debug keynote - Present problems and future solutions
Bug debug keynote - Present problems and future solutions
 
How Four Statistical Rules Forecast Who Wins a Competitive Bid
How Four Statistical Rules Forecast Who Wins a Competitive BidHow Four Statistical Rules Forecast Who Wins a Competitive Bid
How Four Statistical Rules Forecast Who Wins a Competitive Bid
 
Engineering Data Science Objectives for Social Network Analysis
Engineering Data Science Objectives for Social Network AnalysisEngineering Data Science Objectives for Social Network Analysis
Engineering Data Science Objectives for Social Network Analysis
 
The Art of Identifying Vulnerabilities - CascadiaFest 2015
The Art of Identifying Vulnerabilities  - CascadiaFest 2015The Art of Identifying Vulnerabilities  - CascadiaFest 2015
The Art of Identifying Vulnerabilities - CascadiaFest 2015
 
Causal inference-for-profit | Dan McKinley | DN18
Causal inference-for-profit | Dan McKinley | DN18Causal inference-for-profit | Dan McKinley | DN18
Causal inference-for-profit | Dan McKinley | DN18
 
DN18 | A/B Testing: Lessons Learned | Dan McKinley | Mailchimp
DN18 | A/B Testing: Lessons Learned | Dan McKinley | MailchimpDN18 | A/B Testing: Lessons Learned | Dan McKinley | Mailchimp
DN18 | A/B Testing: Lessons Learned | Dan McKinley | Mailchimp
 
Cheminformatics Software Development: Case Studies
Cheminformatics Software Development: Case StudiesCheminformatics Software Development: Case Studies
Cheminformatics Software Development: Case Studies
 
University Course Timetabling by using Multi Objective Genetic Algortihms
University Course Timetabling by using Multi Objective Genetic AlgortihmsUniversity Course Timetabling by using Multi Objective Genetic Algortihms
University Course Timetabling by using Multi Objective Genetic Algortihms
 
Lightning Talks: An Innovation Showcase
Lightning Talks: An Innovation ShowcaseLightning Talks: An Innovation Showcase
Lightning Talks: An Innovation Showcase
 
Bringing Red vs. Blue to Machine Learning
Bringing Red vs. Blue to Machine LearningBringing Red vs. Blue to Machine Learning
Bringing Red vs. Blue to Machine Learning
 
lamoureux
lamoureuxlamoureux
lamoureux
 
IntroML_6_Classification_part3
IntroML_6_Classification_part3IntroML_6_Classification_part3
IntroML_6_Classification_part3
 
IntroML_1_Introduction_Tagged.pdf
IntroML_1_Introduction_Tagged.pdfIntroML_1_Introduction_Tagged.pdf
IntroML_1_Introduction_Tagged.pdf
 
IntroML_1_Introduction
IntroML_1_IntroductionIntroML_1_Introduction
IntroML_1_Introduction
 
Greg Wilson - We Know (but ignore) More Than We Think
Greg Wilson - We Know (but ignore) More Than We ThinkGreg Wilson - We Know (but ignore) More Than We Think
Greg Wilson - We Know (but ignore) More Than We Think
 
Scale your Testing and Quality with Automation Engineering and ML - Carlos Ki...
Scale your Testing and Quality with Automation Engineering and ML - Carlos Ki...Scale your Testing and Quality with Automation Engineering and ML - Carlos Ki...
Scale your Testing and Quality with Automation Engineering and ML - Carlos Ki...
 
NeurIPS2023 Keynote: The Many Faces of Responsible AI.pdf
NeurIPS2023 Keynote: The Many Faces of Responsible AI.pdfNeurIPS2023 Keynote: The Many Faces of Responsible AI.pdf
NeurIPS2023 Keynote: The Many Faces of Responsible AI.pdf
 
IntroML_4_Classification
IntroML_4_ClassificationIntroML_4_Classification
IntroML_4_Classification
 
IntroML_4_Classification_2
IntroML_4_Classification_2IntroML_4_Classification_2
IntroML_4_Classification_2
 
Why Do Computational Scientists Trust Their So
Why Do Computational Scientists Trust Their SoWhy Do Computational Scientists Trust Their So
Why Do Computational Scientists Trust Their So
 

More from DivyanshGupta922023

Git Basics walkthough to all basic concept and commands of git
Git Basics walkthough to all basic concept and commands of gitGit Basics walkthough to all basic concept and commands of git
Git Basics walkthough to all basic concept and commands of git
DivyanshGupta922023
 
DHC Microbiome Presentation 4-23-19.pptx
DHC Microbiome Presentation 4-23-19.pptxDHC Microbiome Presentation 4-23-19.pptx
DHC Microbiome Presentation 4-23-19.pptx
DivyanshGupta922023
 
Introduction to Directed Acyclic Graphs.pptx
Introduction to Directed Acyclic Graphs.pptxIntroduction to Directed Acyclic Graphs.pptx
Introduction to Directed Acyclic Graphs.pptx
DivyanshGupta922023
 

More from DivyanshGupta922023 (17)

(Public) FedCM BlinkOn 16 fedcm and privacy sandbox apis
(Public) FedCM BlinkOn 16 fedcm and privacy sandbox apis(Public) FedCM BlinkOn 16 fedcm and privacy sandbox apis
(Public) FedCM BlinkOn 16 fedcm and privacy sandbox apis
 
DevOps The Buzzword - everything about devops
DevOps The Buzzword - everything about devopsDevOps The Buzzword - everything about devops
DevOps The Buzzword - everything about devops
 
Git Basics walkthough to all basic concept and commands of git
Git Basics walkthough to all basic concept and commands of gitGit Basics walkthough to all basic concept and commands of git
Git Basics walkthough to all basic concept and commands of git
 
jquery summit presentation for large scale javascript applications
jquery summit  presentation for large scale javascript applicationsjquery summit  presentation for large scale javascript applications
jquery summit presentation for large scale javascript applications
 
Next.js - ReactPlayIO.pptx
Next.js - ReactPlayIO.pptxNext.js - ReactPlayIO.pptx
Next.js - ReactPlayIO.pptx
 
Management+team.pptx
Management+team.pptxManagement+team.pptx
Management+team.pptx
 
DHC Microbiome Presentation 4-23-19.pptx
DHC Microbiome Presentation 4-23-19.pptxDHC Microbiome Presentation 4-23-19.pptx
DHC Microbiome Presentation 4-23-19.pptx
 
developer-burnout.pdf
developer-burnout.pdfdeveloper-burnout.pdf
developer-burnout.pdf
 
AzureIntro.pptx
AzureIntro.pptxAzureIntro.pptx
AzureIntro.pptx
 
api-driven-development.pdf
api-driven-development.pdfapi-driven-development.pdf
api-driven-development.pdf
 
Internet of Things.pptx
Internet of Things.pptxInternet of Things.pptx
Internet of Things.pptx
 
Functional JS+ ES6.pptx
Functional JS+ ES6.pptxFunctional JS+ ES6.pptx
Functional JS+ ES6.pptx
 
10-security-concepts-lightning-talk 1of2.pptx
10-security-concepts-lightning-talk 1of2.pptx10-security-concepts-lightning-talk 1of2.pptx
10-security-concepts-lightning-talk 1of2.pptx
 
Introduction to Directed Acyclic Graphs.pptx
Introduction to Directed Acyclic Graphs.pptxIntroduction to Directed Acyclic Graphs.pptx
Introduction to Directed Acyclic Graphs.pptx
 
ReactJS presentation.pptx
ReactJS presentation.pptxReactJS presentation.pptx
ReactJS presentation.pptx
 
01-React js Intro.pptx
01-React js Intro.pptx01-React js Intro.pptx
01-React js Intro.pptx
 
Nextjs13.pptx
Nextjs13.pptxNextjs13.pptx
Nextjs13.pptx
 

Recently uploaded

➥🔝 7737669865 🔝▻ mehsana Call-girls in Women Seeking Men 🔝mehsana🔝 Escorts...
➥🔝 7737669865 🔝▻ mehsana Call-girls in Women Seeking Men  🔝mehsana🔝   Escorts...➥🔝 7737669865 🔝▻ mehsana Call-girls in Women Seeking Men  🔝mehsana🔝   Escorts...
➥🔝 7737669865 🔝▻ mehsana Call-girls in Women Seeking Men 🔝mehsana🔝 Escorts...
nirzagarg
 
Call Girls in Prashant Vihar, Delhi 💯 Call Us 🔝9953056974 🔝 Escort Service
Call Girls in Prashant Vihar, Delhi 💯 Call Us 🔝9953056974 🔝 Escort ServiceCall Girls in Prashant Vihar, Delhi 💯 Call Us 🔝9953056974 🔝 Escort Service
Call Girls in Prashant Vihar, Delhi 💯 Call Us 🔝9953056974 🔝 Escort Service
9953056974 Low Rate Call Girls In Saket, Delhi NCR
 
Low Sexy Call Girls In Mohali 9053900678 🥵Have Save And Good Place 🥵
Low Sexy Call Girls In Mohali 9053900678 🥵Have Save And Good Place 🥵Low Sexy Call Girls In Mohali 9053900678 🥵Have Save And Good Place 🥵
Low Sexy Call Girls In Mohali 9053900678 🥵Have Save And Good Place 🥵
Chandigarh Call girls 9053900678 Call girls in Chandigarh
 
Thalassery Escorts Service ☎️ 6378878445 ( Sakshi Sinha ) High Profile Call G...
Thalassery Escorts Service ☎️ 6378878445 ( Sakshi Sinha ) High Profile Call G...Thalassery Escorts Service ☎️ 6378878445 ( Sakshi Sinha ) High Profile Call G...
Thalassery Escorts Service ☎️ 6378878445 ( Sakshi Sinha ) High Profile Call G...
Call Girls In Delhi Whatsup 9873940964 Enjoy Unlimited Pleasure
 
6.High Profile Call Girls In Punjab +919053900678 Punjab Call GirlHigh Profil...
6.High Profile Call Girls In Punjab +919053900678 Punjab Call GirlHigh Profil...6.High Profile Call Girls In Punjab +919053900678 Punjab Call GirlHigh Profil...
6.High Profile Call Girls In Punjab +919053900678 Punjab Call GirlHigh Profil...
@Chandigarh #call #Girls 9053900678 @Call #Girls in @Punjab 9053900678
 
Lucknow ❤CALL GIRL 88759*99948 ❤CALL GIRLS IN Lucknow ESCORT SERVICE❤CALL GIRL
Lucknow ❤CALL GIRL 88759*99948 ❤CALL GIRLS IN Lucknow ESCORT SERVICE❤CALL GIRLLucknow ❤CALL GIRL 88759*99948 ❤CALL GIRLS IN Lucknow ESCORT SERVICE❤CALL GIRL
Lucknow ❤CALL GIRL 88759*99948 ❤CALL GIRLS IN Lucknow ESCORT SERVICE❤CALL GIRL
imonikaupta
 

Recently uploaded (20)

VVIP Pune Call Girls Sinhagad WhatSapp Number 8005736733 With Elite Staff And...
VVIP Pune Call Girls Sinhagad WhatSapp Number 8005736733 With Elite Staff And...VVIP Pune Call Girls Sinhagad WhatSapp Number 8005736733 With Elite Staff And...
VVIP Pune Call Girls Sinhagad WhatSapp Number 8005736733 With Elite Staff And...
 
Call Girls Sangvi Call Me 7737669865 Budget Friendly No Advance BookingCall G...
Call Girls Sangvi Call Me 7737669865 Budget Friendly No Advance BookingCall G...Call Girls Sangvi Call Me 7737669865 Budget Friendly No Advance BookingCall G...
Call Girls Sangvi Call Me 7737669865 Budget Friendly No Advance BookingCall G...
 
Nanded City ( Call Girls ) Pune 6297143586 Hot Model With Sexy Bhabi Ready ...
Nanded City ( Call Girls ) Pune  6297143586  Hot Model With Sexy Bhabi Ready ...Nanded City ( Call Girls ) Pune  6297143586  Hot Model With Sexy Bhabi Ready ...
Nanded City ( Call Girls ) Pune 6297143586 Hot Model With Sexy Bhabi Ready ...
 
WhatsApp 📞 8448380779 ✅Call Girls In Mamura Sector 66 ( Noida)
WhatsApp 📞 8448380779 ✅Call Girls In Mamura Sector 66 ( Noida)WhatsApp 📞 8448380779 ✅Call Girls In Mamura Sector 66 ( Noida)
WhatsApp 📞 8448380779 ✅Call Girls In Mamura Sector 66 ( Noida)
 
➥🔝 7737669865 🔝▻ mehsana Call-girls in Women Seeking Men 🔝mehsana🔝 Escorts...
➥🔝 7737669865 🔝▻ mehsana Call-girls in Women Seeking Men  🔝mehsana🔝   Escorts...➥🔝 7737669865 🔝▻ mehsana Call-girls in Women Seeking Men  🔝mehsana🔝   Escorts...
➥🔝 7737669865 🔝▻ mehsana Call-girls in Women Seeking Men 🔝mehsana🔝 Escorts...
 
Wadgaon Sheri $ Call Girls Pune 10k @ I'm VIP Independent Escorts Girls 80057...
Wadgaon Sheri $ Call Girls Pune 10k @ I'm VIP Independent Escorts Girls 80057...Wadgaon Sheri $ Call Girls Pune 10k @ I'm VIP Independent Escorts Girls 80057...
Wadgaon Sheri $ Call Girls Pune 10k @ I'm VIP Independent Escorts Girls 80057...
 
20240509 QFM015 Engineering Leadership Reading List April 2024.pdf
20240509 QFM015 Engineering Leadership Reading List April 2024.pdf20240509 QFM015 Engineering Leadership Reading List April 2024.pdf
20240509 QFM015 Engineering Leadership Reading List April 2024.pdf
 
𓀤Call On 7877925207 𓀤 Ahmedguda Call Girls Hot Model With Sexy Bhabi Ready Fo...
𓀤Call On 7877925207 𓀤 Ahmedguda Call Girls Hot Model With Sexy Bhabi Ready Fo...𓀤Call On 7877925207 𓀤 Ahmedguda Call Girls Hot Model With Sexy Bhabi Ready Fo...
𓀤Call On 7877925207 𓀤 Ahmedguda Call Girls Hot Model With Sexy Bhabi Ready Fo...
 
Ganeshkhind ! Call Girls Pune - 450+ Call Girl Cash Payment 8005736733 Neha T...
Ganeshkhind ! Call Girls Pune - 450+ Call Girl Cash Payment 8005736733 Neha T...Ganeshkhind ! Call Girls Pune - 450+ Call Girl Cash Payment 8005736733 Neha T...
Ganeshkhind ! Call Girls Pune - 450+ Call Girl Cash Payment 8005736733 Neha T...
 
Call Girls in Prashant Vihar, Delhi 💯 Call Us 🔝9953056974 🔝 Escort Service
Call Girls in Prashant Vihar, Delhi 💯 Call Us 🔝9953056974 🔝 Escort ServiceCall Girls in Prashant Vihar, Delhi 💯 Call Us 🔝9953056974 🔝 Escort Service
Call Girls in Prashant Vihar, Delhi 💯 Call Us 🔝9953056974 🔝 Escort Service
 
Low Sexy Call Girls In Mohali 9053900678 🥵Have Save And Good Place 🥵
Low Sexy Call Girls In Mohali 9053900678 🥵Have Save And Good Place 🥵Low Sexy Call Girls In Mohali 9053900678 🥵Have Save And Good Place 🥵
Low Sexy Call Girls In Mohali 9053900678 🥵Have Save And Good Place 🥵
 
Thalassery Escorts Service ☎️ 6378878445 ( Sakshi Sinha ) High Profile Call G...
Thalassery Escorts Service ☎️ 6378878445 ( Sakshi Sinha ) High Profile Call G...Thalassery Escorts Service ☎️ 6378878445 ( Sakshi Sinha ) High Profile Call G...
Thalassery Escorts Service ☎️ 6378878445 ( Sakshi Sinha ) High Profile Call G...
 
APNIC Policy Roundup, presented by Sunny Chendi at the 5th ICANN APAC-TWNIC E...
APNIC Policy Roundup, presented by Sunny Chendi at the 5th ICANN APAC-TWNIC E...APNIC Policy Roundup, presented by Sunny Chendi at the 5th ICANN APAC-TWNIC E...
APNIC Policy Roundup, presented by Sunny Chendi at the 5th ICANN APAC-TWNIC E...
 
6.High Profile Call Girls In Punjab +919053900678 Punjab Call GirlHigh Profil...
6.High Profile Call Girls In Punjab +919053900678 Punjab Call GirlHigh Profil...6.High Profile Call Girls In Punjab +919053900678 Punjab Call GirlHigh Profil...
6.High Profile Call Girls In Punjab +919053900678 Punjab Call GirlHigh Profil...
 
Dubai=Desi Dubai Call Girls O525547819 Outdoor Call Girls Dubai
Dubai=Desi Dubai Call Girls O525547819 Outdoor Call Girls DubaiDubai=Desi Dubai Call Girls O525547819 Outdoor Call Girls Dubai
Dubai=Desi Dubai Call Girls O525547819 Outdoor Call Girls Dubai
 
Real Men Wear Diapers T Shirts sweatshirt
Real Men Wear Diapers T Shirts sweatshirtReal Men Wear Diapers T Shirts sweatshirt
Real Men Wear Diapers T Shirts sweatshirt
 
APNIC Updates presented by Paul Wilson at ARIN 53
APNIC Updates presented by Paul Wilson at ARIN 53APNIC Updates presented by Paul Wilson at ARIN 53
APNIC Updates presented by Paul Wilson at ARIN 53
 
Call Girls Ludhiana Just Call 98765-12871 Top Class Call Girl Service Available
Call Girls Ludhiana Just Call 98765-12871 Top Class Call Girl Service AvailableCall Girls Ludhiana Just Call 98765-12871 Top Class Call Girl Service Available
Call Girls Ludhiana Just Call 98765-12871 Top Class Call Girl Service Available
 
Lucknow ❤CALL GIRL 88759*99948 ❤CALL GIRLS IN Lucknow ESCORT SERVICE❤CALL GIRL
Lucknow ❤CALL GIRL 88759*99948 ❤CALL GIRLS IN Lucknow ESCORT SERVICE❤CALL GIRLLucknow ❤CALL GIRL 88759*99948 ❤CALL GIRLS IN Lucknow ESCORT SERVICE❤CALL GIRL
Lucknow ❤CALL GIRL 88759*99948 ❤CALL GIRLS IN Lucknow ESCORT SERVICE❤CALL GIRL
 
All Time Service Available Call Girls Mg Road 👌 ⏭️ 6378878445
All Time Service Available Call Girls Mg Road 👌 ⏭️ 6378878445All Time Service Available Call Girls Mg Road 👌 ⏭️ 6378878445
All Time Service Available Call Girls Mg Road 👌 ⏭️ 6378878445
 

AAAI19-Open.pptx

  • 1. LEARNING AND THE UNKNOWN: SURVEYING STEPS TOWARD OPEN WORLD RECOGNITION Terrance E. Boult IEEE Fellow El Pomar Prof. of Innovation and Security University of Colorado Colorado Springs I’ll post video of this talk at https://github.com/vastab
  • 2. Bachelor of InnovationTM University of Colorado Colorado Springs WHAT “CLASS” IS THIS OBJECT?  A few will actually know it  Most will (quickly) think “I don’t know” – you know you don’t know which is OpenSet  The curious might try to look it up and get more data to “learn” it’s a Ctenophore aka a comb jelly – i.e. we do open world learning
  • 3. Bachelor of InnovationTM University of Colorado Colorado Springs TRAFFICKCAM EXAMPLE (AAAI19 PAPER) HOTELS-50K: A GLOBAL HOTEL RECOGNITION DATASET: ABBY STYLIANOU, HONG XUAN, MAYA SHENDE, JONATHAN BRANDT, RICHARD SOUVENIR, ROBERT PLESS Adversaries will try to defeat a system
  • 4. Bachelor of InnovationTM University of Colorado Colorado Springs SURVEILLANCE EXAMPLE: CAR DETECTION/COUNTING Nature can produce persistent/long-lived unknown inputs e.g. Ice on camera housing. (Also bug, bird doodoo…)
  • 5. Bachelor of InnovationTM University of Colorado Colorado Springs Gallery Probe Match Score =0.72705 Face Verification (mid JANUS IARPA program)
  • 6. Bachelor of InnovationTM University of Colorado Colorado Springs Match Score= 0.99769 Gallery Probe Face Verification (mid-JANUS IARPA program, top performer) L2-Softmax trained Cosine-distance matching
  • 7. Bachelor of InnovationTM University of Colorado Colorado Springs WHAT WENT WRONG? Absence of Evidence is not Evidence of Absence Being far from boundaries & training evidence, A of E, implies high “probabilities” in classifiers such as SVM or Softmax The open set/world is full of “unknowns” that will absent in training! Bayesian Reasoning cannot help us as we cannot normalize without the probability of the unknown inputs
  • 8. Bachelor of InnovationTM University of Colorado Colorado Springs The Range of Openness/Unknowns in problems Multi-class Classification Face Verification Detection Open S et Recognition Closed Open ? ? ? ? ? ? ? ? ? ? Training and testing samples come from known classes Claimed identity, possibility for impostors One class, everything else in the world is negative Multiple known classes,many unknown classes Multi-class Classification Face Verification Detection Open S et Recognition Closed Open ? ? ? ? ? ? ? ? ? ? Training and testing samples come from known classes Claimed identity, possibility for impostors One class, everything else in the world is negative Multiple known classes,many unknown classes Paper has > 90 citations covering OSR from 14 different application areas Multi-class Classification Face Verification Detection Open S et Recognition Closed Open ? ? ? ? ? ? ? ? ? ? Training and testing samples come from known classes Claimed identity, possibility for impostors One class, everything else in the world is negative Multiple known classes,many unknown classes
  • 9. Bachelor of InnovationTM University of Colorado Colorado Springs Thresholding “Probability” vs Open Set On optimum recognition error and reject tradeoff. C. Chow, IEEE Trans. Info. Theory 16(1):41–46. 1970
  • 10. Bachelor of InnovationTM University of Colorado Colorado Springs LEARNING IN THE FACE OF UNKNOWN UNKNOWNS: FORMALIZATION OF OPEN-SET RECOGNITION Open Space Risk Empirical Risk/Error Scheirer et al. TPAMI ‘13 𝑉 is set of Valid Class training samples; 𝐾 is set of Known unknowns (backgrounds)
  • 11. Bachelor of InnovationTM University of Colorado Colorado Springs OPEN SPACE RISK “open space” is the space far from known samples. A simple risk model a constant penalty for labeling that anything other than unknown in a ratio such as:
  • 12. Bachelor of InnovationTM University of Colorado Colorado Springs Algorithms that solve OSR RBF SVM WHAT SOLVES OSR? Any detector that uses pure linear classifiers, linear SVM, HAAR cascades, or Softmax-based classifiers, will almost always have an unbounded open set risk, and hence does not solve OSR even with thresholding. GMM SVDD If they include a “bias” or Bayesian normalization they probably don’t solve OSR. WSVM PI-SVM KDE EVM NNO
  • 13. Bachelor of InnovationTM University of Colorado Colorado Springs “NOVELTY DETECTION, ANOMALY DETECTION AND DETECTING “OUT OF DISTRIBUTION SAMPLE” Long history and many many papers on the first 2, while the latter is a new term for similar ideas or learning with outliers. Open-Set Recognition ≅ Anomaly/Novelty detection + Multi-class Recognition Compute Novelty or Anomaly score Is Outlier ? ”Closed set” multi-class Label Does NOT directly address “open set recognition” but can be used in sequence to address OSR. However, rarely is open set recognition part of evaluation evaluation.
  • 14. Bachelor of InnovationTM University of Colorado Colorado Springs  Classic machine learning presumes all classes known and classifies all of feature space. Thresholding vs Open Set
  • 15. Bachelor of InnovationTM University of Colorado Colorado Springs COMPACT ABATING PROBABILITY W-SVM ~= OneClassRBF * (EVT scaled) Binary RBF SVM
  • 16. Bachelor of InnovationTM University of Colorado Colorado Springs THERE ARE MANY PROBLEM VARIANTS AND SOLUTIONS  It can be a part of zero/few shot learning  Xian, Y.; Lampert, C. H.; Schiele, B.; and Akata, Z. 2018. Zero- shot learning-A comprehensive evaluation of the good, the bad and the ugly. IEEE TPAMI 2018.  Or Open set clustering/incremental with no labels  Active Sampling for Open-Set Classification without Initial Annotation Z-Y. Liu and S.-J Huang AAA19 (Tech Session 1: Weakly Supervised Learning 1 Thu 2-3:30, Coral Ballroom 3-5)  Problems where you have to predict scores on unseen data, e.g.  “An Open-World Extension to Knowledge Graph Completion Models” H. Shah et al AAAI 19 (Tech Session 3: Thu 10:25-11:25, Coral 2)
  • 17. Bachelor of InnovationTM University of Colorado Colorado Springs • World with Knowns (K) & • Known Unknowns (KU) Unknowns Unknowns (UU) OSR: Recognizes known classes or Detect as Unknown • NU: Novel Unknowns Collect & Label Data • LU: Labeled Unknowns Incremental Class Learning OpenSet Network Feature Training Class Label Unknown Bendale-Boult CVPR15 Open World Learning
  • 18. Bachelor of InnovationTM University of Colorado Colorado Springs EXTREME VALUE MACHINE  Uses our “margin distribution theorem” to derive EVT-based non-linear models that are provably Open-set and also can do Open World/”Incremental”. Can use classic or deep features Rudd Et Al. TPAMI 18
  • 19. Bachelor of InnovationTM University of Colorado Colorado Springs TOWARDS OPEN-SET DEEP NETWORKS Abhijit Bendale*, Terrance Boult Samsung Research America* University of Colorado of Colorado Springs
  • 20. Bachelor of InnovationTM University of Colorado Colorado Springs OpenMax for Deep Networks Distance from MAV Frequency FC7 FC8 AlexNet OpenMax CAP Model using using EVT On distances from MAV
  • 21. Bachelor of InnovationTM University of Colorado Colorado Springs OPEN-SET DEEP NETWORKS Model (MAV) Real Image Softmax: 0.94, baseball Fooling Image Softmax: 1.0, baseball Open-Set Image Softmax :0.15, baseball Openmax : 0.94, baseball Openmax: 0.00, baseball 0.95 Unknown Openmax: 0.17, baseball : 0.80, Unknown Model (MAV) Real Image Fooling Image Open-Set Image
  • 22. Bachelor of InnovationTM University of Colorado Colorado Springs OPENMAX ONLY SOMEWHAT BETTER. WE RESEARCHED WHY.  Lenet++ 2D feature representation
  • 23. Bachelor of InnovationTM University of Colorado Colorado Springs TRAINING/TESTING  Known Unknowns NIST Letters Known: MNIST
  • 24. Bachelor of InnovationTM University of Colorado Colorado Springs LENET++ trained on MNIST MNIST test set (Colors) LENET++ trained on MNIST MNIST test set (colors) Black = NIST Letters.. Rather than being far away or “outside” the data, features for unknown inputs generally overlap known classes
  • 25. Bachelor of InnovationTM University of Colorado Colorado Springs OBSERVATION FROM DEFAULT RESPONSE – LEADING TO OUR APPROACH Observer that there is difference in entropy and magnitude. While Open-Set limited response outside the ring of data, most of the unknowns had smaller magnitude. The NeurIPS18 approach seeks to emphasize that difference.
  • 26. Bachelor of InnovationTM University of Colorado Colorado Springs GOAL OF OUR APPROACH
  • 27. Bachelor of InnovationTM University of Colorado Colorado Springs LENET++ RESPONSES TO KNOWNS AND UNKNOWNS. Colored dots represent test samples from the ten MNIST classes, while black dots represent samples from unknown unknowns. The dashed gray-white lines indicate class borders. The figures in the bottom are histograms of network scores for known (green) and unknowns (red) with logarithmic vertical axis. Dhamija et al. 18
  • 28. Bachelor of InnovationTM University of Colorado Colorado Springs Dhamija et al. 18
  • 29. Bachelor of InnovationTM University of Colorado Colorado Springs AISTATS ‘11
  • 30. Bachelor of InnovationTM University of Colorado Colorado Springs Matching Deep Features from VGG2 set at FAR=10-4 says this pic of TB matches Barack Obama So does a commercial system Adversarial Examples show we do NOT understand how deep network actually work– “Close in input is not close in features”. Until we do understand we cannot really do open world deep networks.
  • 31. Bachelor of InnovationTM University of Colorado Colorado Springs CONCLUSIONS  We cannot anticipate and train for all “unknown inputs”  Bayesian reasoning cannot help us if we don’t know probability of ”unknown” inputs occurring.  Almost all classical classifiers have unbounded risk and make highly confident errors. OSR tools address both.  Traditional deep network map unknown ontop of knowns  Starting to make progress on deep networks, but it’s a area with lots of research potential. Adversarial Examples show there is major issues still “unknown.
  • 32. Bachelor of InnovationTM University of Colorado Colorado Springs “Intelligence comes with hard work and curiosity for the unknown.” Roberto Llamas Do not fear the unknown — join us in taming it. https://github.com/vastab Our code is mostly LIBSVM or BSD-3 ”free” licensed.

Editor's Notes

  1. The first part of the talk will explore issues with  deep networks dealing with  "unknowns" inputs, and the general issues of Open-Set recognition in deep networks.  We review our first attempt at Open-Set deep networks,  "OpenMax," and discuss is successes and limitations and why classic "open-set" approaches don't really solve the problem of deep unknowns.  We then present our ongoing work, to first appear at NIPS2018, on a new model we call the ObjectoSphere.  Using ObjectoSphere loss begins to address the learning of deep features that can handle unknown inputs.  We present examples of its use first on simple datasets sets (MNIST/CFAR) and then on a real-world problem of open-set face recognition.   We then move to another type of unknown for deep networks:  adversarial examples, images  perturbations that are invisible to humans but easily fool deep networks.  While Open-Set recognition tries to deal with inputs that are "far" from known training samples, these adversarial examples are in perceptually close in input space but far in feature face.  This last part of the talk will discuss various potential theories about the causes of adversarial examples,  why we know those theories are not correct, and why they show we don't understand deep networks.  We introduce our deep-feature adversarial approach, called LOTS, and return to the examples of object-recognition and face-recognition showing how our LOTS adversarial examples can successfully attack even open-set recognition systems.
  2. Why does this happen? We believe, the closed set nature of deep networks forces them to choose from one of the known classes leading to such artifacts. Our hypothesis is that deep networks suffer from the open space, prone to any discriminative classifier. The softmax layer divides the output space into N half spaces each with potentially infinite volume. However, recognition in the real world is Open-Set, i.e. the recognition system should reject unknown/unseen classes at test time.
  3. Gaussian Mixture Modules, kernel density estimators, RBF SVMs or Support Vector Data Descriptors (SVDD), may, but do not have to have bounded open space risk. It depends how they are combined and thresholded. If they include a “bias” they probably don’t solve OSR.
  4. Hi. I am Abhijit Bendale and I will present work on Towards Open-Set Deep Networks. This is joint work with Prof. Terrance Boult at University of Colorado.
  5. To explore the approaches we used LeNet++ which a has a 2D feature space just before the SoftMax classifier.
  6. Observations leading to our NeuralIPS 18 spotlight paper
  7. Performance of two new losses (red and green) are significantly better than any prior approach.