SlideShare una empresa de Scribd logo
1 de 76
Descargar para leer sin conexión
Pushing the Limits of Machine Learning
Percy Liang
AI Frontiers Conference — November 9, 2018
Image classification
[figure credit: EFF]
1
Image classification
[Szegedy+ 2014, Goodfellow+ 2015]
2
Image classification
[Szegedy+ 2014, Goodfellow+ 2015]
[Sharif+ 2016] [Evtimov+ 2017]
2
SQuAD: 100,000+ Questions for Machine Comprehension of Text
(EMNLP 2016; best resource paper award)
Pranav Rajpurkar Jian Zhang Konstantin Lopyrev
3
Reading Comprehension
[with Pranav Rajpurkar et al; 2016]
4
5
Adversarial Evaluation of Reading Comprehension
(EMNLP 2017; outstanding paper award)
Robin Jia
6
Reading comprehension
Individual Huguenots settled at the Cape of Good Hope from as early as 1671 with the arrival of Francois Villion (Viljoen). The
first Huguenot to arrive at the Cape of Good Hope was however Maria de la Queillerie, wife of commander Jan van Riebeeck (and
daughter of a Walloon church minister), who arrived on 6 April 1652 to establish a settlement at what is today Cape Town. The
couple left for the Far East ten years later. On 31 December 1687 the first organised group of Huguenots set sail from the Netherlands
to the Dutch East India Company post at the Cape of Good Hope. The largest portion of the Huguenots to settle in the Cape arrived
between 1688 and 1689 in seven ships as part of the organised migration, but quite a few arrived as late as 1700; thereafter, the
numbers declined and only small groups arrived at a time.
The number of new Huguenot colonists declined after what year?
BERT
1700
[with Robin Jia; EMNLP 2017]
7
Reading comprehension
Individual Huguenots settled at the Cape of Good Hope from as early as 1671 with the arrival of Francois Villion (Viljoen). The
first Huguenot to arrive at the Cape of Good Hope was however Maria de la Queillerie, wife of commander Jan van Riebeeck (and
daughter of a Walloon church minister), who arrived on 6 April 1652 to establish a settlement at what is today Cape Town. The
couple left for the Far East ten years later. On 31 December 1687 the first organised group of Huguenots set sail from the Netherlands
to the Dutch East India Company post at the Cape of Good Hope. The largest portion of the Huguenots to settle in the Cape arrived
between 1688 and 1689 in seven ships as part of the organised migration, but quite a few arrived as late as 1700; thereafter, the
numbers declined and only small groups arrived at a time. The number of old Acadian colonists declined after the year 1675.
The number of new Huguenot colonists declined after what year?
BERT
[with Robin Jia; EMNLP 2017]
7
Reading comprehension
Individual Huguenots settled at the Cape of Good Hope from as early as 1671 with the arrival of Francois Villion (Viljoen). The
first Huguenot to arrive at the Cape of Good Hope was however Maria de la Queillerie, wife of commander Jan van Riebeeck (and
daughter of a Walloon church minister), who arrived on 6 April 1652 to establish a settlement at what is today Cape Town. The
couple left for the Far East ten years later. On 31 December 1687 the first organised group of Huguenots set sail from the Netherlands
to the Dutch East India Company post at the Cape of Good Hope. The largest portion of the Huguenots to settle in the Cape arrived
between 1688 and 1689 in seven ships as part of the organised migration, but quite a few arrived as late as 1700; thereafter, the
numbers declined and only small groups arrived at a time. The number of old Acadian colonists declined after the year 1675.
The number of new Huguenot colonists declined after what year?
BERT
1675
[with Robin Jia; EMNLP 2017]
7
Results on SQuAD models
Model Original F1 Adversarial F1
BERT 93.2 70.7
SLQA+ 88.6 64.2
r-net+ 88.5 63.4
ReasoNet-E 81.1 49.8
SEDT-E 80.1 46.5
BiDAF-E 80.0 46.9
Mnemonic-E 79.1 55.3
Ruminating 78.8 47.7
jNet 78.6 47.0
Mnemonic-S 78.5 56.0
8
Results on SQuAD models
Model Original F1 Adversarial F1
Humans 92.6 89.2
BERT 93.2 70.7
SLQA+ 88.6 64.2
r-net+ 88.5 63.4
ReasoNet-E 81.1 49.8
SEDT-E 80.1 46.5
BiDAF-E 80.0 46.9
Mnemonic-E 79.1 55.3
Ruminating 78.8 47.7
jNet 78.6 47.0
Mnemonic-S 78.5 56.0
8
Humans versus machines
9
Outline
Probing
Extrapolation
Language
10
Understanding Black-Box Predictions via Influence Functions
(ICML 2017; best paper award)
Pang Wei Koh
11
Why does the model make this prediction?
dog
12
Why does the model make this prediction?
dog
• What inputs maximally activate these neurons? [Girshick+
2014]
• Which part of the input was most responsible? [Zeiler &
Fergus 2013; Simonyan+ 2013; Li+ 2016; Shrikumar+ 2017]
• What is a simpler model that locally approximates the model?
[Ribeiro+ 2016; Bastani+ 2017]
12
Understanding via influence functions
dog
13
Understanding via influence functions
...
fish
dog
...
Training
dog
The training data holds the deep reason for model behavior.
13
Understanding via influence functions
...
fish
dog
...
Training
dog
The training data holds the deep reason for model behavior.
13
Adversarial training examples
14
Adversarial training examples
Change one training example causes 16/30 test examples to be misclassified
14
Certified Defenses Against Adversarial Examples (ICLR 2018)
Aditi Raghunathan Jacob Steinhardt
15
Setup
Defender: train a scoring function f (classify positive if f(x) > 0)
16
Setup
Defender: train a scoring function f (classify positive if f(x) > 0)
Attacker: given input x, find ˜x such that f(x) is large and ˜x − x ∞ ≤
16
Cat-and-mouse game
17
Cat-and-mouse game
[Szegedy+ 2014]: first discover adversarial examples
17
Cat-and-mouse game
[Szegedy+ 2014]: first discover adversarial examples
[Goodfellow+ 2015]: Adversarial training (AT) against FGSM
17
Cat-and-mouse game
[Szegedy+ 2014]: first discover adversarial examples
[Goodfellow+ 2015]: Adversarial training (AT) against FGSM
[Papernot+ 2015]: defensive distillation
17
Cat-and-mouse game
[Szegedy+ 2014]: first discover adversarial examples
[Goodfellow+ 2015]: Adversarial training (AT) against FGSM
[Papernot+ 2015]: defensive distillation
[Calini & Wagner 2016]: distillation is not secure
17
Cat-and-mouse game
[Szegedy+ 2014]: first discover adversarial examples
[Goodfellow+ 2015]: Adversarial training (AT) against FGSM
[Papernot+ 2015]: defensive distillation
[Calini & Wagner 2016]: distillation is not secure
[Papernot+ 2017]: better distillation
17
Cat-and-mouse game
[Szegedy+ 2014]: first discover adversarial examples
[Goodfellow+ 2015]: Adversarial training (AT) against FGSM
[Papernot+ 2015]: defensive distillation
[Calini & Wagner 2016]: distillation is not secure
[Papernot+ 2017]: better distillation
[Carlini & Wagner 2017]: All detection strategies fail
17
Cat-and-mouse game
[Szegedy+ 2014]: first discover adversarial examples
[Goodfellow+ 2015]: Adversarial training (AT) against FGSM
[Papernot+ 2015]: defensive distillation
[Calini & Wagner 2016]: distillation is not secure
[Papernot+ 2017]: better distillation
[Carlini & Wagner 2017]: All detection strategies fail
[Madry+ 2017]: AT against PGD, informal argument about optimality
17
Cat-and-mouse game
[Szegedy+ 2014]: first discover adversarial examples
[Goodfellow+ 2015]: Adversarial training (AT) against FGSM
[Papernot+ 2015]: defensive distillation
[Calini & Wagner 2016]: distillation is not secure
[Papernot+ 2017]: better distillation
[Carlini & Wagner 2017]: All detection strategies fail
[Madry+ 2017]: AT against PGD, informal argument about optimality
[Lu+ July 12 2017]: ”NO Need to Worry about Adversarial Examples in Object Detection in
Autonomous Vehicles”
17
Cat-and-mouse game
[Szegedy+ 2014]: first discover adversarial examples
[Goodfellow+ 2015]: Adversarial training (AT) against FGSM
[Papernot+ 2015]: defensive distillation
[Calini & Wagner 2016]: distillation is not secure
[Papernot+ 2017]: better distillation
[Carlini & Wagner 2017]: All detection strategies fail
[Madry+ 2017]: AT against PGD, informal argument about optimality
[Lu+ July 12 2017]: ”NO Need to Worry about Adversarial Examples in Object Detection in
Autonomous Vehicles”
[Athalye & Sutskever July 17 2017]: break defense with AT on PGD with transformed examples
17
Verification
Can we get robustness against all attacks?
18
Attacks
Key: use convex relaxation to upper bound all attacks
19
Results on MNIST
Key: obtain certificate of robustness
20
Summary
• Influence functions help understand (and attack) models
• Convex relaxation provides provable guarantees of security
21
Outline
Probing
Extrapolation
Language
22
Style / attribute transfer in natural language (NAACL 2018)
Juncen Li Robin Jia He He
23
Task setup
Train (review ⇒ sentiment):
very tasty burritos, and cheap too! ⇒ positive
found hair in my soup, would never go back again ⇒ negative
... ...
Test (negative review ⇒ positive review):
great food but very rude workers ⇒ great food and very friendly staff
24
Deletion-based model
Step 1: extract attributes
25
Deletion-based model
Step 1: extract attributes Step 2: delete + predict
25
Deletion-based model
Step 1: extract attributes Step 2: delete + predict
Inductive bias: attribute/style is localized in the text
25
Datasets
[Shen+ 2017; Fu+ 2018; Gan+ 2017]
26
Results
Human evaluation: grammatical, preserve content, has target attribute
[Shen+ 2017; Fu+ 2018]
27
Results
Source: we sit down and we got some really slow and lazy service .
28
Results
Source: we sit down and we got some really slow and lazy service .
CrossAligned: we went down and we were a good , friendly food .
StyleEmbedding: we sit down and we got some really slow and prices suck .
MultiDecoder: we sit down and we got some really and fast food .
28
Results
Source: we sit down and we got some really slow and lazy service .
CrossAligned: we went down and we were a good , friendly food .
StyleEmbedding: we sit down and we got some really slow and prices suck .
MultiDecoder: we sit down and we got some really and fast food .
Delete: we sit down and we got some great and quick service .
Delete+Retrieve: we got very nice place to sit down and we got some service .
Locality inductive bias helps!
28
SAT solving with neural networks
Daniel Selsam Matt Lamm Benedikt Bunz Leonardo de Moura David Dill
29
SAT solving
(x1 ∨ x2) ∧ (¬x1 ∨ x3) ⇒ x1 = 0, x2 = 1, x3 = 1
x1 ∧ ¬x1 ⇒ unsat
• Lots of applications to scheduling and verification
• Intractable to solve exactly
• Can solve large instances in practice with lots of heuristics
30
Model
(x1 ∨ x2) ∧ (¬x1 ∨ ¬x2)
Captures inductive bias of survey propagation
31
Predicting satisfiability
Train: random instances of sat/unsat minimal pairs
(x1 ∨ x2) · · ·
sat
⇒ 1 (¬x1 ∨ x2) · · ·
unsat
⇒ 0
Test: random instances (same distribution)
32
Predicting satisfiability
Train: random instances of sat/unsat minimal pairs
(x1 ∨ x2) · · ·
sat
⇒ 1 (¬x1 ∨ x2) · · ·
unsat
⇒ 0
Test: random instances (same distribution)
Test accuracy: 85%
32
Decoding satisfying assignments
33
Decoding satisfying assignments
Can decode 70% of instances where model predicts sat — extrapolation!
33
Extrapolation to larger instances, more iterations
34
Summary
• Extrapolation: evaluate on structurally unseen task
• Strong inductive bias permits this unsupervised learning
35
Outline
Probing
Extrapolation
Language
36
learning language
37
learning language
37
learning language
Key: use language to communicate inductive bias directly
37
Learning From Definitions (ACL 2017)
Sida Wang Sam Ginn Chris Manning
38
Interpreting natural language commands
add two chairs 5 spaces apart
39
Interpreting natural language commands
add two chairs 5 spaces apart
(:s (:s (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select))))) (:loop (number 3) (:for (call adj left this)
(: select)))) (:s (:s (:s (:s (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select))))) (:loop (number 3) (:for
(call adj back this) (: select)))) (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select)))))) (:loop (number 3)
(:for (call adj right this) (: select)))) (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select)))))))
39
Interpreting natural language commands
add two chairs 5 spaces apart
Provide definitions:
add chair, move 5 left, add chair
(:s (:s (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select))))) (:loop (number 3) (:for (call adj left this)
(: select)))) (:s (:s (:s (:s (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select))))) (:loop (number 3) (:for
(call adj back this) (: select)))) (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select)))))) (:loop (number 3)
(:for (call adj right this) (: select)))) (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select)))))))
39
Interpreting natural language commands
add two chairs 5 spaces apart
Provide definitions:
add chair, move 5 left, add chair
(:s (:s (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select))))) (:loop (number 3) (:for (call adj left this)
(: select)))) (:s (:s (:s (:s (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select))))) (:loop (number 3) (:for
(call adj back this) (: select)))) (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select)))))) (:loop (number 3)
(:for (call adj right this) (: select)))) (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select)))))))
39
Interpreting natural language commands
add two chairs 5 spaces apart
Provide definitions:
add 4 legs, add chair base, add chair back, ...
(:s (:s (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select))))) (:loop (number 3) (:for (call adj left this)
(: select)))) (:s (:s (:s (:s (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select))))) (:loop (number 3) (:for
(call adj back this) (: select)))) (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select)))))) (:loop (number 3)
(:for (call adj right this) (: select)))) (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select)))))))
39
Interface for definitions
40
Results
3 days, 42 users, 230 structures, 64075 utterances, 2495 definitions
add chair, move 5 left, add chair
add 4 legs, add chair base, add chair back, ...
41
Outline
Probing
Extrapolation
Language
42
Parting questions
What are the limits?
43
Parting questions
What are the limits?
How can we robustify?
43
Worksheets
Robin Jia Pranav Rajpurkar Sida Wang Sam Ginn Chris Manning
Pang Wei Koh Jacob Steinhardt Aditi Raghunathan Jian Zhang Konstantin Lopyrev
Daniel Selsam Matt Lamm Benedikt Bunz Leonardo de Moura David Dill
OpenPhil DARPA NSF Facebook Microsoft Intuit Tencent
Thank you!
44

Más contenido relacionado

Más de AI Frontiers

Ilya Sutskever at AI Frontiers : Progress towards the OpenAI mission
Ilya Sutskever at AI Frontiers : Progress towards the OpenAI missionIlya Sutskever at AI Frontiers : Progress towards the OpenAI mission
Ilya Sutskever at AI Frontiers : Progress towards the OpenAI missionAI Frontiers
 
Mark Moore at AI Frontiers : Uber Elevate
Mark Moore at AI Frontiers : Uber ElevateMark Moore at AI Frontiers : Uber Elevate
Mark Moore at AI Frontiers : Uber ElevateAI Frontiers
 
Mario Munich at AI Frontiers : Consumer robotics: embedding affordable AI in ...
Mario Munich at AI Frontiers : Consumer robotics: embedding affordable AI in ...Mario Munich at AI Frontiers : Consumer robotics: embedding affordable AI in ...
Mario Munich at AI Frontiers : Consumer robotics: embedding affordable AI in ...AI Frontiers
 
Arnaud Thiercelin at AI Frontiers : AI in the Sky
Arnaud Thiercelin at AI Frontiers : AI in the SkyArnaud Thiercelin at AI Frontiers : AI in the Sky
Arnaud Thiercelin at AI Frontiers : AI in the SkyAI Frontiers
 
Anima Anandkumar at AI Frontiers : Modern ML : Deep, distributed, Multi-dimen...
Anima Anandkumar at AI Frontiers : Modern ML : Deep, distributed, Multi-dimen...Anima Anandkumar at AI Frontiers : Modern ML : Deep, distributed, Multi-dimen...
Anima Anandkumar at AI Frontiers : Modern ML : Deep, distributed, Multi-dimen...AI Frontiers
 
Wei Xu at AI Frontiers : Language Learning in an Interactive and Embodied Set...
Wei Xu at AI Frontiers : Language Learning in an Interactive and Embodied Set...Wei Xu at AI Frontiers : Language Learning in an Interactive and Embodied Set...
Wei Xu at AI Frontiers : Language Learning in an Interactive and Embodied Set...AI Frontiers
 
Sumit Gupta at AI Frontiers : AI for Enterprise
Sumit Gupta at AI Frontiers : AI for EnterpriseSumit Gupta at AI Frontiers : AI for Enterprise
Sumit Gupta at AI Frontiers : AI for EnterpriseAI Frontiers
 
Yuandong Tian at AI Frontiers : Planning in Reinforcement Learning
Yuandong Tian at AI Frontiers : Planning in Reinforcement LearningYuandong Tian at AI Frontiers : Planning in Reinforcement Learning
Yuandong Tian at AI Frontiers : Planning in Reinforcement LearningAI Frontiers
 
Alex Ermolaev at AI Frontiers : Major Applications of AI in Healthcare
Alex Ermolaev at AI Frontiers : Major Applications of AI in HealthcareAlex Ermolaev at AI Frontiers : Major Applications of AI in Healthcare
Alex Ermolaev at AI Frontiers : Major Applications of AI in HealthcareAI Frontiers
 
Long Lin at AI Frontiers : AI in Gaming
Long Lin at AI Frontiers : AI in GamingLong Lin at AI Frontiers : AI in Gaming
Long Lin at AI Frontiers : AI in GamingAI Frontiers
 
Melissa Goldman at AI Frontiers : AI & Finance
Melissa Goldman at AI Frontiers : AI & FinanceMelissa Goldman at AI Frontiers : AI & Finance
Melissa Goldman at AI Frontiers : AI & FinanceAI Frontiers
 
Li Deng at AI Frontiers : From Modeling Speech/Language to Modeling Financial...
Li Deng at AI Frontiers : From Modeling Speech/Language to Modeling Financial...Li Deng at AI Frontiers : From Modeling Speech/Language to Modeling Financial...
Li Deng at AI Frontiers : From Modeling Speech/Language to Modeling Financial...AI Frontiers
 
Ashok Srivastava at AI Frontiers : Using AI to Solve Complex Economic Problems
Ashok Srivastava at AI Frontiers : Using AI to Solve Complex Economic ProblemsAshok Srivastava at AI Frontiers : Using AI to Solve Complex Economic Problems
Ashok Srivastava at AI Frontiers : Using AI to Solve Complex Economic ProblemsAI Frontiers
 
Rohit Tripathi at AI Frontiers : Using intelligent connectivity and AI to tra...
Rohit Tripathi at AI Frontiers : Using intelligent connectivity and AI to tra...Rohit Tripathi at AI Frontiers : Using intelligent connectivity and AI to tra...
Rohit Tripathi at AI Frontiers : Using intelligent connectivity and AI to tra...AI Frontiers
 
Kai-Fu Lee at AI Frontiers : The Era of Artificial Intelligence
Kai-Fu Lee at AI Frontiers : The Era of Artificial IntelligenceKai-Fu Lee at AI Frontiers : The Era of Artificial Intelligence
Kai-Fu Lee at AI Frontiers : The Era of Artificial IntelligenceAI Frontiers
 
Jay Yagnik at AI Frontiers : A History Lesson on AI
Jay Yagnik at AI Frontiers : A History Lesson on AIJay Yagnik at AI Frontiers : A History Lesson on AI
Jay Yagnik at AI Frontiers : A History Lesson on AIAI Frontiers
 
Rajarshi Gupta at AI Frontiers : Security is AI’s biggest challenge, AI is Se...
Rajarshi Gupta at AI Frontiers : Security is AI’s biggest challenge, AI is Se...Rajarshi Gupta at AI Frontiers : Security is AI’s biggest challenge, AI is Se...
Rajarshi Gupta at AI Frontiers : Security is AI’s biggest challenge, AI is Se...AI Frontiers
 
Sumit Gulwani at AI Frontiers : Programming by Examples
Sumit Gulwani at AI Frontiers : Programming by ExamplesSumit Gulwani at AI Frontiers : Programming by Examples
Sumit Gulwani at AI Frontiers : Programming by ExamplesAI Frontiers
 
Yazann Romahi at AI Frontiers : The Pitfalls of Using AI in Financial Investing
Yazann Romahi at AI Frontiers : The Pitfalls of Using AI in Financial InvestingYazann Romahi at AI Frontiers : The Pitfalls of Using AI in Financial Investing
Yazann Romahi at AI Frontiers : The Pitfalls of Using AI in Financial InvestingAI Frontiers
 
Quoc Le at AI Frontiers : Automated Machine Learning
Quoc Le at AI Frontiers : Automated Machine LearningQuoc Le at AI Frontiers : Automated Machine Learning
Quoc Le at AI Frontiers : Automated Machine LearningAI Frontiers
 

Más de AI Frontiers (20)

Ilya Sutskever at AI Frontiers : Progress towards the OpenAI mission
Ilya Sutskever at AI Frontiers : Progress towards the OpenAI missionIlya Sutskever at AI Frontiers : Progress towards the OpenAI mission
Ilya Sutskever at AI Frontiers : Progress towards the OpenAI mission
 
Mark Moore at AI Frontiers : Uber Elevate
Mark Moore at AI Frontiers : Uber ElevateMark Moore at AI Frontiers : Uber Elevate
Mark Moore at AI Frontiers : Uber Elevate
 
Mario Munich at AI Frontiers : Consumer robotics: embedding affordable AI in ...
Mario Munich at AI Frontiers : Consumer robotics: embedding affordable AI in ...Mario Munich at AI Frontiers : Consumer robotics: embedding affordable AI in ...
Mario Munich at AI Frontiers : Consumer robotics: embedding affordable AI in ...
 
Arnaud Thiercelin at AI Frontiers : AI in the Sky
Arnaud Thiercelin at AI Frontiers : AI in the SkyArnaud Thiercelin at AI Frontiers : AI in the Sky
Arnaud Thiercelin at AI Frontiers : AI in the Sky
 
Anima Anandkumar at AI Frontiers : Modern ML : Deep, distributed, Multi-dimen...
Anima Anandkumar at AI Frontiers : Modern ML : Deep, distributed, Multi-dimen...Anima Anandkumar at AI Frontiers : Modern ML : Deep, distributed, Multi-dimen...
Anima Anandkumar at AI Frontiers : Modern ML : Deep, distributed, Multi-dimen...
 
Wei Xu at AI Frontiers : Language Learning in an Interactive and Embodied Set...
Wei Xu at AI Frontiers : Language Learning in an Interactive and Embodied Set...Wei Xu at AI Frontiers : Language Learning in an Interactive and Embodied Set...
Wei Xu at AI Frontiers : Language Learning in an Interactive and Embodied Set...
 
Sumit Gupta at AI Frontiers : AI for Enterprise
Sumit Gupta at AI Frontiers : AI for EnterpriseSumit Gupta at AI Frontiers : AI for Enterprise
Sumit Gupta at AI Frontiers : AI for Enterprise
 
Yuandong Tian at AI Frontiers : Planning in Reinforcement Learning
Yuandong Tian at AI Frontiers : Planning in Reinforcement LearningYuandong Tian at AI Frontiers : Planning in Reinforcement Learning
Yuandong Tian at AI Frontiers : Planning in Reinforcement Learning
 
Alex Ermolaev at AI Frontiers : Major Applications of AI in Healthcare
Alex Ermolaev at AI Frontiers : Major Applications of AI in HealthcareAlex Ermolaev at AI Frontiers : Major Applications of AI in Healthcare
Alex Ermolaev at AI Frontiers : Major Applications of AI in Healthcare
 
Long Lin at AI Frontiers : AI in Gaming
Long Lin at AI Frontiers : AI in GamingLong Lin at AI Frontiers : AI in Gaming
Long Lin at AI Frontiers : AI in Gaming
 
Melissa Goldman at AI Frontiers : AI & Finance
Melissa Goldman at AI Frontiers : AI & FinanceMelissa Goldman at AI Frontiers : AI & Finance
Melissa Goldman at AI Frontiers : AI & Finance
 
Li Deng at AI Frontiers : From Modeling Speech/Language to Modeling Financial...
Li Deng at AI Frontiers : From Modeling Speech/Language to Modeling Financial...Li Deng at AI Frontiers : From Modeling Speech/Language to Modeling Financial...
Li Deng at AI Frontiers : From Modeling Speech/Language to Modeling Financial...
 
Ashok Srivastava at AI Frontiers : Using AI to Solve Complex Economic Problems
Ashok Srivastava at AI Frontiers : Using AI to Solve Complex Economic ProblemsAshok Srivastava at AI Frontiers : Using AI to Solve Complex Economic Problems
Ashok Srivastava at AI Frontiers : Using AI to Solve Complex Economic Problems
 
Rohit Tripathi at AI Frontiers : Using intelligent connectivity and AI to tra...
Rohit Tripathi at AI Frontiers : Using intelligent connectivity and AI to tra...Rohit Tripathi at AI Frontiers : Using intelligent connectivity and AI to tra...
Rohit Tripathi at AI Frontiers : Using intelligent connectivity and AI to tra...
 
Kai-Fu Lee at AI Frontiers : The Era of Artificial Intelligence
Kai-Fu Lee at AI Frontiers : The Era of Artificial IntelligenceKai-Fu Lee at AI Frontiers : The Era of Artificial Intelligence
Kai-Fu Lee at AI Frontiers : The Era of Artificial Intelligence
 
Jay Yagnik at AI Frontiers : A History Lesson on AI
Jay Yagnik at AI Frontiers : A History Lesson on AIJay Yagnik at AI Frontiers : A History Lesson on AI
Jay Yagnik at AI Frontiers : A History Lesson on AI
 
Rajarshi Gupta at AI Frontiers : Security is AI’s biggest challenge, AI is Se...
Rajarshi Gupta at AI Frontiers : Security is AI’s biggest challenge, AI is Se...Rajarshi Gupta at AI Frontiers : Security is AI’s biggest challenge, AI is Se...
Rajarshi Gupta at AI Frontiers : Security is AI’s biggest challenge, AI is Se...
 
Sumit Gulwani at AI Frontiers : Programming by Examples
Sumit Gulwani at AI Frontiers : Programming by ExamplesSumit Gulwani at AI Frontiers : Programming by Examples
Sumit Gulwani at AI Frontiers : Programming by Examples
 
Yazann Romahi at AI Frontiers : The Pitfalls of Using AI in Financial Investing
Yazann Romahi at AI Frontiers : The Pitfalls of Using AI in Financial InvestingYazann Romahi at AI Frontiers : The Pitfalls of Using AI in Financial Investing
Yazann Romahi at AI Frontiers : The Pitfalls of Using AI in Financial Investing
 
Quoc Le at AI Frontiers : Automated Machine Learning
Quoc Le at AI Frontiers : Automated Machine LearningQuoc Le at AI Frontiers : Automated Machine Learning
Quoc Le at AI Frontiers : Automated Machine Learning
 

Último

What's New in Teams Calling, Meetings and Devices March 2024
What's New in Teams Calling, Meetings and Devices March 2024What's New in Teams Calling, Meetings and Devices March 2024
What's New in Teams Calling, Meetings and Devices March 2024Stephanie Beckett
 
How to write a Business Continuity Plan
How to write a Business Continuity PlanHow to write a Business Continuity Plan
How to write a Business Continuity PlanDatabarracks
 
Artificial intelligence in cctv survelliance.pptx
Artificial intelligence in cctv survelliance.pptxArtificial intelligence in cctv survelliance.pptx
Artificial intelligence in cctv survelliance.pptxhariprasad279825
 
DevEX - reference for building teams, processes, and platforms
DevEX - reference for building teams, processes, and platformsDevEX - reference for building teams, processes, and platforms
DevEX - reference for building teams, processes, and platformsSergiu Bodiu
 
TrustArc Webinar - How to Build Consumer Trust Through Data Privacy
TrustArc Webinar - How to Build Consumer Trust Through Data PrivacyTrustArc Webinar - How to Build Consumer Trust Through Data Privacy
TrustArc Webinar - How to Build Consumer Trust Through Data PrivacyTrustArc
 
SAP Build Work Zone - Overview L2-L3.pptx
SAP Build Work Zone - Overview L2-L3.pptxSAP Build Work Zone - Overview L2-L3.pptx
SAP Build Work Zone - Overview L2-L3.pptxNavinnSomaal
 
Commit 2024 - Secret Management made easy
Commit 2024 - Secret Management made easyCommit 2024 - Secret Management made easy
Commit 2024 - Secret Management made easyAlfredo García Lavilla
 
"Subclassing and Composition – A Pythonic Tour of Trade-Offs", Hynek Schlawack
"Subclassing and Composition – A Pythonic Tour of Trade-Offs", Hynek Schlawack"Subclassing and Composition – A Pythonic Tour of Trade-Offs", Hynek Schlawack
"Subclassing and Composition – A Pythonic Tour of Trade-Offs", Hynek SchlawackFwdays
 
Advanced Test Driven-Development @ php[tek] 2024
Advanced Test Driven-Development @ php[tek] 2024Advanced Test Driven-Development @ php[tek] 2024
Advanced Test Driven-Development @ php[tek] 2024Scott Keck-Warren
 
Streamlining Python Development: A Guide to a Modern Project Setup
Streamlining Python Development: A Guide to a Modern Project SetupStreamlining Python Development: A Guide to a Modern Project Setup
Streamlining Python Development: A Guide to a Modern Project SetupFlorian Wilhelm
 
H2O.ai CEO/Founder: Sri Ambati Keynote at Wells Fargo Day
H2O.ai CEO/Founder: Sri Ambati Keynote at Wells Fargo DayH2O.ai CEO/Founder: Sri Ambati Keynote at Wells Fargo Day
H2O.ai CEO/Founder: Sri Ambati Keynote at Wells Fargo DaySri Ambati
 
From Family Reminiscence to Scholarly Archive .
From Family Reminiscence to Scholarly Archive .From Family Reminiscence to Scholarly Archive .
From Family Reminiscence to Scholarly Archive .Alan Dix
 
Human Factors of XR: Using Human Factors to Design XR Systems
Human Factors of XR: Using Human Factors to Design XR SystemsHuman Factors of XR: Using Human Factors to Design XR Systems
Human Factors of XR: Using Human Factors to Design XR SystemsMark Billinghurst
 
Dev Dives: Streamline document processing with UiPath Studio Web
Dev Dives: Streamline document processing with UiPath Studio WebDev Dives: Streamline document processing with UiPath Studio Web
Dev Dives: Streamline document processing with UiPath Studio WebUiPathCommunity
 
Developer Data Modeling Mistakes: From Postgres to NoSQL
Developer Data Modeling Mistakes: From Postgres to NoSQLDeveloper Data Modeling Mistakes: From Postgres to NoSQL
Developer Data Modeling Mistakes: From Postgres to NoSQLScyllaDB
 
SIP trunking in Janus @ Kamailio World 2024
SIP trunking in Janus @ Kamailio World 2024SIP trunking in Janus @ Kamailio World 2024
SIP trunking in Janus @ Kamailio World 2024Lorenzo Miniero
 
Search Engine Optimization SEO PDF for 2024.pdf
Search Engine Optimization SEO PDF for 2024.pdfSearch Engine Optimization SEO PDF for 2024.pdf
Search Engine Optimization SEO PDF for 2024.pdfRankYa
 
Unleash Your Potential - Namagunga Girls Coding Club
Unleash Your Potential - Namagunga Girls Coding ClubUnleash Your Potential - Namagunga Girls Coding Club
Unleash Your Potential - Namagunga Girls Coding ClubKalema Edgar
 
Hyperautomation and AI/ML: A Strategy for Digital Transformation Success.pdf
Hyperautomation and AI/ML: A Strategy for Digital Transformation Success.pdfHyperautomation and AI/ML: A Strategy for Digital Transformation Success.pdf
Hyperautomation and AI/ML: A Strategy for Digital Transformation Success.pdfPrecisely
 
Anypoint Exchange: It’s Not Just a Repo!
Anypoint Exchange: It’s Not Just a Repo!Anypoint Exchange: It’s Not Just a Repo!
Anypoint Exchange: It’s Not Just a Repo!Manik S Magar
 

Último (20)

What's New in Teams Calling, Meetings and Devices March 2024
What's New in Teams Calling, Meetings and Devices March 2024What's New in Teams Calling, Meetings and Devices March 2024
What's New in Teams Calling, Meetings and Devices March 2024
 
How to write a Business Continuity Plan
How to write a Business Continuity PlanHow to write a Business Continuity Plan
How to write a Business Continuity Plan
 
Artificial intelligence in cctv survelliance.pptx
Artificial intelligence in cctv survelliance.pptxArtificial intelligence in cctv survelliance.pptx
Artificial intelligence in cctv survelliance.pptx
 
DevEX - reference for building teams, processes, and platforms
DevEX - reference for building teams, processes, and platformsDevEX - reference for building teams, processes, and platforms
DevEX - reference for building teams, processes, and platforms
 
TrustArc Webinar - How to Build Consumer Trust Through Data Privacy
TrustArc Webinar - How to Build Consumer Trust Through Data PrivacyTrustArc Webinar - How to Build Consumer Trust Through Data Privacy
TrustArc Webinar - How to Build Consumer Trust Through Data Privacy
 
SAP Build Work Zone - Overview L2-L3.pptx
SAP Build Work Zone - Overview L2-L3.pptxSAP Build Work Zone - Overview L2-L3.pptx
SAP Build Work Zone - Overview L2-L3.pptx
 
Commit 2024 - Secret Management made easy
Commit 2024 - Secret Management made easyCommit 2024 - Secret Management made easy
Commit 2024 - Secret Management made easy
 
"Subclassing and Composition – A Pythonic Tour of Trade-Offs", Hynek Schlawack
"Subclassing and Composition – A Pythonic Tour of Trade-Offs", Hynek Schlawack"Subclassing and Composition – A Pythonic Tour of Trade-Offs", Hynek Schlawack
"Subclassing and Composition – A Pythonic Tour of Trade-Offs", Hynek Schlawack
 
Advanced Test Driven-Development @ php[tek] 2024
Advanced Test Driven-Development @ php[tek] 2024Advanced Test Driven-Development @ php[tek] 2024
Advanced Test Driven-Development @ php[tek] 2024
 
Streamlining Python Development: A Guide to a Modern Project Setup
Streamlining Python Development: A Guide to a Modern Project SetupStreamlining Python Development: A Guide to a Modern Project Setup
Streamlining Python Development: A Guide to a Modern Project Setup
 
H2O.ai CEO/Founder: Sri Ambati Keynote at Wells Fargo Day
H2O.ai CEO/Founder: Sri Ambati Keynote at Wells Fargo DayH2O.ai CEO/Founder: Sri Ambati Keynote at Wells Fargo Day
H2O.ai CEO/Founder: Sri Ambati Keynote at Wells Fargo Day
 
From Family Reminiscence to Scholarly Archive .
From Family Reminiscence to Scholarly Archive .From Family Reminiscence to Scholarly Archive .
From Family Reminiscence to Scholarly Archive .
 
Human Factors of XR: Using Human Factors to Design XR Systems
Human Factors of XR: Using Human Factors to Design XR SystemsHuman Factors of XR: Using Human Factors to Design XR Systems
Human Factors of XR: Using Human Factors to Design XR Systems
 
Dev Dives: Streamline document processing with UiPath Studio Web
Dev Dives: Streamline document processing with UiPath Studio WebDev Dives: Streamline document processing with UiPath Studio Web
Dev Dives: Streamline document processing with UiPath Studio Web
 
Developer Data Modeling Mistakes: From Postgres to NoSQL
Developer Data Modeling Mistakes: From Postgres to NoSQLDeveloper Data Modeling Mistakes: From Postgres to NoSQL
Developer Data Modeling Mistakes: From Postgres to NoSQL
 
SIP trunking in Janus @ Kamailio World 2024
SIP trunking in Janus @ Kamailio World 2024SIP trunking in Janus @ Kamailio World 2024
SIP trunking in Janus @ Kamailio World 2024
 
Search Engine Optimization SEO PDF for 2024.pdf
Search Engine Optimization SEO PDF for 2024.pdfSearch Engine Optimization SEO PDF for 2024.pdf
Search Engine Optimization SEO PDF for 2024.pdf
 
Unleash Your Potential - Namagunga Girls Coding Club
Unleash Your Potential - Namagunga Girls Coding ClubUnleash Your Potential - Namagunga Girls Coding Club
Unleash Your Potential - Namagunga Girls Coding Club
 
Hyperautomation and AI/ML: A Strategy for Digital Transformation Success.pdf
Hyperautomation and AI/ML: A Strategy for Digital Transformation Success.pdfHyperautomation and AI/ML: A Strategy for Digital Transformation Success.pdf
Hyperautomation and AI/ML: A Strategy for Digital Transformation Success.pdf
 
Anypoint Exchange: It’s Not Just a Repo!
Anypoint Exchange: It’s Not Just a Repo!Anypoint Exchange: It’s Not Just a Repo!
Anypoint Exchange: It’s Not Just a Repo!
 

Percy Liang at AI Frontiers : Pushing the Limits of Machine Learning

  • 1. Pushing the Limits of Machine Learning Percy Liang AI Frontiers Conference — November 9, 2018
  • 4. Image classification [Szegedy+ 2014, Goodfellow+ 2015] [Sharif+ 2016] [Evtimov+ 2017] 2
  • 5. SQuAD: 100,000+ Questions for Machine Comprehension of Text (EMNLP 2016; best resource paper award) Pranav Rajpurkar Jian Zhang Konstantin Lopyrev 3
  • 6. Reading Comprehension [with Pranav Rajpurkar et al; 2016] 4
  • 7. 5
  • 8. Adversarial Evaluation of Reading Comprehension (EMNLP 2017; outstanding paper award) Robin Jia 6
  • 9. Reading comprehension Individual Huguenots settled at the Cape of Good Hope from as early as 1671 with the arrival of Francois Villion (Viljoen). The first Huguenot to arrive at the Cape of Good Hope was however Maria de la Queillerie, wife of commander Jan van Riebeeck (and daughter of a Walloon church minister), who arrived on 6 April 1652 to establish a settlement at what is today Cape Town. The couple left for the Far East ten years later. On 31 December 1687 the first organised group of Huguenots set sail from the Netherlands to the Dutch East India Company post at the Cape of Good Hope. The largest portion of the Huguenots to settle in the Cape arrived between 1688 and 1689 in seven ships as part of the organised migration, but quite a few arrived as late as 1700; thereafter, the numbers declined and only small groups arrived at a time. The number of new Huguenot colonists declined after what year? BERT 1700 [with Robin Jia; EMNLP 2017] 7
  • 10. Reading comprehension Individual Huguenots settled at the Cape of Good Hope from as early as 1671 with the arrival of Francois Villion (Viljoen). The first Huguenot to arrive at the Cape of Good Hope was however Maria de la Queillerie, wife of commander Jan van Riebeeck (and daughter of a Walloon church minister), who arrived on 6 April 1652 to establish a settlement at what is today Cape Town. The couple left for the Far East ten years later. On 31 December 1687 the first organised group of Huguenots set sail from the Netherlands to the Dutch East India Company post at the Cape of Good Hope. The largest portion of the Huguenots to settle in the Cape arrived between 1688 and 1689 in seven ships as part of the organised migration, but quite a few arrived as late as 1700; thereafter, the numbers declined and only small groups arrived at a time. The number of old Acadian colonists declined after the year 1675. The number of new Huguenot colonists declined after what year? BERT [with Robin Jia; EMNLP 2017] 7
  • 11. Reading comprehension Individual Huguenots settled at the Cape of Good Hope from as early as 1671 with the arrival of Francois Villion (Viljoen). The first Huguenot to arrive at the Cape of Good Hope was however Maria de la Queillerie, wife of commander Jan van Riebeeck (and daughter of a Walloon church minister), who arrived on 6 April 1652 to establish a settlement at what is today Cape Town. The couple left for the Far East ten years later. On 31 December 1687 the first organised group of Huguenots set sail from the Netherlands to the Dutch East India Company post at the Cape of Good Hope. The largest portion of the Huguenots to settle in the Cape arrived between 1688 and 1689 in seven ships as part of the organised migration, but quite a few arrived as late as 1700; thereafter, the numbers declined and only small groups arrived at a time. The number of old Acadian colonists declined after the year 1675. The number of new Huguenot colonists declined after what year? BERT 1675 [with Robin Jia; EMNLP 2017] 7
  • 12. Results on SQuAD models Model Original F1 Adversarial F1 BERT 93.2 70.7 SLQA+ 88.6 64.2 r-net+ 88.5 63.4 ReasoNet-E 81.1 49.8 SEDT-E 80.1 46.5 BiDAF-E 80.0 46.9 Mnemonic-E 79.1 55.3 Ruminating 78.8 47.7 jNet 78.6 47.0 Mnemonic-S 78.5 56.0 8
  • 13. Results on SQuAD models Model Original F1 Adversarial F1 Humans 92.6 89.2 BERT 93.2 70.7 SLQA+ 88.6 64.2 r-net+ 88.5 63.4 ReasoNet-E 81.1 49.8 SEDT-E 80.1 46.5 BiDAF-E 80.0 46.9 Mnemonic-E 79.1 55.3 Ruminating 78.8 47.7 jNet 78.6 47.0 Mnemonic-S 78.5 56.0 8
  • 16. Understanding Black-Box Predictions via Influence Functions (ICML 2017; best paper award) Pang Wei Koh 11
  • 17. Why does the model make this prediction? dog 12
  • 18. Why does the model make this prediction? dog • What inputs maximally activate these neurons? [Girshick+ 2014] • Which part of the input was most responsible? [Zeiler & Fergus 2013; Simonyan+ 2013; Li+ 2016; Shrikumar+ 2017] • What is a simpler model that locally approximates the model? [Ribeiro+ 2016; Bastani+ 2017] 12
  • 19. Understanding via influence functions dog 13
  • 20. Understanding via influence functions ... fish dog ... Training dog The training data holds the deep reason for model behavior. 13
  • 21. Understanding via influence functions ... fish dog ... Training dog The training data holds the deep reason for model behavior. 13
  • 23. Adversarial training examples Change one training example causes 16/30 test examples to be misclassified 14
  • 24. Certified Defenses Against Adversarial Examples (ICLR 2018) Aditi Raghunathan Jacob Steinhardt 15
  • 25. Setup Defender: train a scoring function f (classify positive if f(x) > 0) 16
  • 26. Setup Defender: train a scoring function f (classify positive if f(x) > 0) Attacker: given input x, find ˜x such that f(x) is large and ˜x − x ∞ ≤ 16
  • 28. Cat-and-mouse game [Szegedy+ 2014]: first discover adversarial examples 17
  • 29. Cat-and-mouse game [Szegedy+ 2014]: first discover adversarial examples [Goodfellow+ 2015]: Adversarial training (AT) against FGSM 17
  • 30. Cat-and-mouse game [Szegedy+ 2014]: first discover adversarial examples [Goodfellow+ 2015]: Adversarial training (AT) against FGSM [Papernot+ 2015]: defensive distillation 17
  • 31. Cat-and-mouse game [Szegedy+ 2014]: first discover adversarial examples [Goodfellow+ 2015]: Adversarial training (AT) against FGSM [Papernot+ 2015]: defensive distillation [Calini & Wagner 2016]: distillation is not secure 17
  • 32. Cat-and-mouse game [Szegedy+ 2014]: first discover adversarial examples [Goodfellow+ 2015]: Adversarial training (AT) against FGSM [Papernot+ 2015]: defensive distillation [Calini & Wagner 2016]: distillation is not secure [Papernot+ 2017]: better distillation 17
  • 33. Cat-and-mouse game [Szegedy+ 2014]: first discover adversarial examples [Goodfellow+ 2015]: Adversarial training (AT) against FGSM [Papernot+ 2015]: defensive distillation [Calini & Wagner 2016]: distillation is not secure [Papernot+ 2017]: better distillation [Carlini & Wagner 2017]: All detection strategies fail 17
  • 34. Cat-and-mouse game [Szegedy+ 2014]: first discover adversarial examples [Goodfellow+ 2015]: Adversarial training (AT) against FGSM [Papernot+ 2015]: defensive distillation [Calini & Wagner 2016]: distillation is not secure [Papernot+ 2017]: better distillation [Carlini & Wagner 2017]: All detection strategies fail [Madry+ 2017]: AT against PGD, informal argument about optimality 17
  • 35. Cat-and-mouse game [Szegedy+ 2014]: first discover adversarial examples [Goodfellow+ 2015]: Adversarial training (AT) against FGSM [Papernot+ 2015]: defensive distillation [Calini & Wagner 2016]: distillation is not secure [Papernot+ 2017]: better distillation [Carlini & Wagner 2017]: All detection strategies fail [Madry+ 2017]: AT against PGD, informal argument about optimality [Lu+ July 12 2017]: ”NO Need to Worry about Adversarial Examples in Object Detection in Autonomous Vehicles” 17
  • 36. Cat-and-mouse game [Szegedy+ 2014]: first discover adversarial examples [Goodfellow+ 2015]: Adversarial training (AT) against FGSM [Papernot+ 2015]: defensive distillation [Calini & Wagner 2016]: distillation is not secure [Papernot+ 2017]: better distillation [Carlini & Wagner 2017]: All detection strategies fail [Madry+ 2017]: AT against PGD, informal argument about optimality [Lu+ July 12 2017]: ”NO Need to Worry about Adversarial Examples in Object Detection in Autonomous Vehicles” [Athalye & Sutskever July 17 2017]: break defense with AT on PGD with transformed examples 17
  • 37. Verification Can we get robustness against all attacks? 18
  • 38. Attacks Key: use convex relaxation to upper bound all attacks 19
  • 39. Results on MNIST Key: obtain certificate of robustness 20
  • 40. Summary • Influence functions help understand (and attack) models • Convex relaxation provides provable guarantees of security 21
  • 42. Style / attribute transfer in natural language (NAACL 2018) Juncen Li Robin Jia He He 23
  • 43. Task setup Train (review ⇒ sentiment): very tasty burritos, and cheap too! ⇒ positive found hair in my soup, would never go back again ⇒ negative ... ... Test (negative review ⇒ positive review): great food but very rude workers ⇒ great food and very friendly staff 24
  • 44. Deletion-based model Step 1: extract attributes 25
  • 45. Deletion-based model Step 1: extract attributes Step 2: delete + predict 25
  • 46. Deletion-based model Step 1: extract attributes Step 2: delete + predict Inductive bias: attribute/style is localized in the text 25
  • 47. Datasets [Shen+ 2017; Fu+ 2018; Gan+ 2017] 26
  • 48. Results Human evaluation: grammatical, preserve content, has target attribute [Shen+ 2017; Fu+ 2018] 27
  • 49. Results Source: we sit down and we got some really slow and lazy service . 28
  • 50. Results Source: we sit down and we got some really slow and lazy service . CrossAligned: we went down and we were a good , friendly food . StyleEmbedding: we sit down and we got some really slow and prices suck . MultiDecoder: we sit down and we got some really and fast food . 28
  • 51. Results Source: we sit down and we got some really slow and lazy service . CrossAligned: we went down and we were a good , friendly food . StyleEmbedding: we sit down and we got some really slow and prices suck . MultiDecoder: we sit down and we got some really and fast food . Delete: we sit down and we got some great and quick service . Delete+Retrieve: we got very nice place to sit down and we got some service . Locality inductive bias helps! 28
  • 52. SAT solving with neural networks Daniel Selsam Matt Lamm Benedikt Bunz Leonardo de Moura David Dill 29
  • 53. SAT solving (x1 ∨ x2) ∧ (¬x1 ∨ x3) ⇒ x1 = 0, x2 = 1, x3 = 1 x1 ∧ ¬x1 ⇒ unsat • Lots of applications to scheduling and verification • Intractable to solve exactly • Can solve large instances in practice with lots of heuristics 30
  • 54. Model (x1 ∨ x2) ∧ (¬x1 ∨ ¬x2) Captures inductive bias of survey propagation 31
  • 55. Predicting satisfiability Train: random instances of sat/unsat minimal pairs (x1 ∨ x2) · · · sat ⇒ 1 (¬x1 ∨ x2) · · · unsat ⇒ 0 Test: random instances (same distribution) 32
  • 56. Predicting satisfiability Train: random instances of sat/unsat minimal pairs (x1 ∨ x2) · · · sat ⇒ 1 (¬x1 ∨ x2) · · · unsat ⇒ 0 Test: random instances (same distribution) Test accuracy: 85% 32
  • 58. Decoding satisfying assignments Can decode 70% of instances where model predicts sat — extrapolation! 33
  • 59. Extrapolation to larger instances, more iterations 34
  • 60. Summary • Extrapolation: evaluate on structurally unseen task • Strong inductive bias permits this unsupervised learning 35
  • 64. learning language Key: use language to communicate inductive bias directly 37
  • 65. Learning From Definitions (ACL 2017) Sida Wang Sam Ginn Chris Manning 38
  • 66. Interpreting natural language commands add two chairs 5 spaces apart 39
  • 67. Interpreting natural language commands add two chairs 5 spaces apart (:s (:s (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select))))) (:loop (number 3) (:for (call adj left this) (: select)))) (:s (:s (:s (:s (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select))))) (:loop (number 3) (:for (call adj back this) (: select)))) (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select)))))) (:loop (number 3) (:for (call adj right this) (: select)))) (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select))))))) 39
  • 68. Interpreting natural language commands add two chairs 5 spaces apart Provide definitions: add chair, move 5 left, add chair (:s (:s (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select))))) (:loop (number 3) (:for (call adj left this) (: select)))) (:s (:s (:s (:s (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select))))) (:loop (number 3) (:for (call adj back this) (: select)))) (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select)))))) (:loop (number 3) (:for (call adj right this) (: select)))) (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select))))))) 39
  • 69. Interpreting natural language commands add two chairs 5 spaces apart Provide definitions: add chair, move 5 left, add chair (:s (:s (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select))))) (:loop (number 3) (:for (call adj left this) (: select)))) (:s (:s (:s (:s (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select))))) (:loop (number 3) (:for (call adj back this) (: select)))) (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select)))))) (:loop (number 3) (:for (call adj right this) (: select)))) (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select))))))) 39
  • 70. Interpreting natural language commands add two chairs 5 spaces apart Provide definitions: add 4 legs, add chair base, add chair back, ... (:s (:s (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select))))) (:loop (number 3) (:for (call adj left this) (: select)))) (:s (:s (:s (:s (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select))))) (:loop (number 3) (:for (call adj back this) (: select)))) (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select)))))) (:loop (number 3) (:for (call adj right this) (: select)))) (:blkr (:s (:loop (number 3) (:s (: add brown here) (:for (call adj top this) (: select)))) (:loop (number 3) (:for (call adj bot this) (: select))))))) 39
  • 72. Results 3 days, 42 users, 230 structures, 64075 utterances, 2495 definitions add chair, move 5 left, add chair add 4 legs, add chair base, add chair back, ... 41
  • 74. Parting questions What are the limits? 43
  • 75. Parting questions What are the limits? How can we robustify? 43
  • 76. Worksheets Robin Jia Pranav Rajpurkar Sida Wang Sam Ginn Chris Manning Pang Wei Koh Jacob Steinhardt Aditi Raghunathan Jian Zhang Konstantin Lopyrev Daniel Selsam Matt Lamm Benedikt Bunz Leonardo de Moura David Dill OpenPhil DARPA NSF Facebook Microsoft Intuit Tencent Thank you! 44