In recent years, machine learning has undoubtedly been hugely successful in driving progress in AI applications. However, as we will explore in this talk, even state-of-the-art systems have "blind spots" which make them generalize poorly out of domain and render them vulnerable to adversarial examples. We then suggest that more unsupervised learning settings can encourage the development of more robust systems. We show positive results on two tasks: (i) text style and attribute transfer, the task of converting a sentence with one attribute (e.g., sentiment) to one with another; and (ii) solving SAT instances (classical problems requiring logical reasoning) using end-to-end neural networks.
5. SQuAD: 100,000+ Questions for Machine Comprehension of Text
(EMNLP 2016; best resource paper award)
Pranav Rajpurkar Jian Zhang Konstantin Lopyrev
3
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
ďŹrst 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 ďŹrst 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
ďŹrst 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 ďŹrst 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
ďŹrst 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 ďŹrst 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
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
26. Setup
Defender: train a scoring function f (classify positive if f(x) > 0)
Attacker: given input x, ďŹnd Ëx such that f(x) is large and Ëx â x â â¤
16
30. Cat-and-mouse game
[Szegedy+ 2014]: ďŹrst discover adversarial examples
[Goodfellow+ 2015]: Adversarial training (AT) against FGSM
[Papernot+ 2015]: defensive distillation
17
31. Cat-and-mouse game
[Szegedy+ 2014]: ďŹrst 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]: ďŹrst 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]: ďŹrst 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]: ďŹrst 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]: ďŹrst 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]: âNO Need to Worry about Adversarial Examples in Object Detection in
Autonomous Vehiclesâ
17
36. Cat-and-mouse game
[Szegedy+ 2014]: ďŹrst 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]: âNO Need to Worry about Adversarial Examples in Object Detection in
Autonomous Vehiclesâ
[Athalye & Sutskever July 17 2017]: break defense with AT on PGD with transformed examples
17
42. Style / attribute transfer in natural language (NAACL 2018)
Juncen Li Robin Jia He He
23
43. Task setup
Train (review â sentiment):
very tasty burritos, and cheap too! â positive
found hair in my soup, would never go back again â negative
... ...
Test (negative review â positive review):
great food but very rude workers â great food and very friendly staďŹ
24
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) ⧠(ÂŹx1 ⨠x3) â x1 = 0, x2 = 1, x3 = 1
x1 ⧠x1 â unsat
⢠Lots of applications to scheduling and veriďŹcation
⢠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 satisďŹability
Train: random instances of sat/unsat minimal pairs
(x1 ⨠x2) ¡ ¡ ¡
sat
â 1 (ÂŹx1 ⨠x2) ¡ ¡ ¡
unsat
â 0
Test: random instances (same distribution)
32
56. Predicting satisďŹability
Train: random instances of sat/unsat minimal pairs
(x1 ⨠x2) ¡ ¡ ¡
sat
â 1 (ÂŹx1 ⨠x2) ¡ ¡ ¡
unsat
â 0
Test: random instances (same distribution)
Test accuracy: 85%
32
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