Lecture 27: Ethics in computer vision (part 2) 1 Announcements - - PowerPoint PPT Presentation

Lecture 27: Ethics in computer vision (part 2)

1

• Last class :(
• Extra office hours:
• Today: end of class until 2pm
• Thurs: 12:00 - 1:00pm.
• PS8 grades out (regrade requests due Friday)
• PS9, PS10 due tonight. No late days allowed.

Announcements

2

A machine learning algorithm will do whatever the training data tells it to do. If the data is bad or biased, the learned algorithm will be too.

Garbage in, garbage out

3

Source: Isola, Torralba, Freeman

Chatbot released on twitter. Learned from interactions with users 
 
 
 
 
 
 Started mimicking offensive language, was shut down.

Microsoft’s Tay chatbot

Image source: https://money.cnn.com/2016/03/30/technology/tay-tweets-microsoft/index.html

4

The Giraffe-Tree problem

[“Measuring Machine Intelligence Through Visual Question Answering”, Zitnick et al., 2016]

5

Test Train

Nearest neighbor baseline

Source: L. Zitnick

6

A black and white cat sitting in a bathroom sink. Two zebras and a giraffe in a field.

Nearest Neighbor

See mscoco.org for image information

Source: L. Zitnick

7

A man riding a motorcycle on a beach. An airplane is parked on the tarmac at an airport.

Source: L. Zitnick

8

Image captioning

Results

CIDEr-D Meteor ROUGE-L BLEU-4 Google[4] 0.943 0.254 0.53 0.309 MSR Captivator[9] 0.931 0.248 0.526 0.308 m-RNN[15] 0.917 0.242 0.521 0.299 MSR[8] 0.912 0.247 0.519 0.291 Nearest Neighbor[11] 0.886 0.237 0.507 0.280 m-RNN (Baidu/ UCLA)[16] 0.886 0.238 0.524 0.302 Berkeley LRCN[2] 0.869 0.242 0.517 0.277 Human[5] 0.854 0.252 0.484 0.217 Montreal/Toronto[10] 0.85 0.243 0.513 0.268 PicSOM[13] 0.833 0.231 0.505 0.281 MLBL[7] 0.74 0.219 0.499 0.26 ACVT[1] 0.709 0.213 0.483 0.246 NeuralTalk[12] 0.674 0.21 0.475 0.224 Tsinghua Bigeye[14] 0.673 0.207 0.49 0.241 MIL[6] 0.666 0.214 0.468 0.216 Brno University[3] 0.517 0.195 0.403 0.134

COCO Caption Challenge

Source: L. Zitnick

9

Visual Question Answering Dataset

Source: L. Zitnick

[“Colorful image colorization”, Zhang et al., ECCV 2016]

11

Source: Isola, Torralba, Freeman

[“Colorful image colorization”, Zhang et al., ECCV 2016]

12

[“Colorful image colorization”, Zhang et al., ECCV 2016]

13

Source: Isola, Torralba, Freeman

14

Generalization

What Google thinks are student bedrooms

Training data

Test data

15

Source: Isola, Torralba, Freeman

Driving simulator (GTA) Driving in the real world

Training data

Test data

16

Need learning methods that can bridge this domain gap!

Source: Isola, Torralba, Freeman

Revisiting the problem of generalization

17

Training data

{xi, yi}N

i=1

X

Y

18

Source: Isola, Torralba, Freeman

Training data

{xi, yi}N

i=1

X

Y

True data-generating process

Test data

X

Y

pdata

{x, y} ∼ pdata

19

Source: Isola, Torralba, Freeman

Training data

X

Y

True data-generating process

Test data

X

Y

pdata

{x(train)

i

, y(train)

i

}

iid

∼ pdata

{x(test)

i

, y(test)

i

}

iid

∼ pdata

{x(train)

i

, y(train)

i

}N

i=1

{x(test)

i

, y(test)

i

}M

i=1

20

Source: Isola, Torralba, Freeman

Training data

X

Y

This is a huge assumption! Almost never true in practice!

Test data

X

Y

{x(train)

i

, y(train)

i

}

iid

∼ pdata

{x(test)

i

, y(test)

i

}

iid

∼ pdata

{x(train)

i

, y(train)

i

}N

i=1

{x(test)

i

, y(test)

i

}M

i=1

21

Source: Isola, Torralba, Freeman

Training data

X

Y

Much more commonly, we have

Test data

X

Y

{x(train)

i

, y(train)

i

}N

i=1

{x(test)

i

, y(test)

i

}M

i=1

{x(test)

i

, y(test)

i

}

iid

∼ ptest

{x(train)

i

, y(train)

i

}

iid

∼ ptrain

ptrain 6= ptest

22

Source: Isola, Torralba, Freeman

Training data

X

Y

Test data

X

Y

{x(train)

i

, y(train)

i

}N

i=1

{x(test)

i

, y(test)

i

}M

i=1

Our training data did cover the part of the distribution that was tested (biased data) generalization error

23

Source: Isola, Torralba, Freeman

Domain gap between and will cause us to fail to generalize.

ptrain

ptest

training domain testing domain

(where we actual use our model) Space of natural images Training data Test data

24

Source: Isola, Torralba, Freeman

https://www.nytimes.com/2018/02/09/technology/facial-recognition-race-artificial-intelligence.html

Social consequences

25

Source: Isola, Torralba, Freeman

Algorithmic Bias

http://proceedings.mlr.press/v81/buolamwini18a/buolamwini18a.pdf

http://gendershades.org/overview.html

26

Source: Isola, Torralba, Freeman

27

https://www.nytimes.com/2017/10/09/science/stanford-sexual-orientation-study.html

Questionable data choices

How can we collect good data?

+ Correctly labeled + Unbiased (good coverage of all relevant kinds of data)

28

Source: Isola, Torralba, Freeman

Crowdsourcing

29

The value of data

The Large Hadron Collider $ 10 10 Amazon Mechanical Turk $ 10 2 - 10 4

30

Source: Isola, Torralba, Freeman

But can humans collect good data?

31

32

Source: Isola, Torralba, Freeman

Labeling to get a Ph.D. Labeling for money


(Sorokin, Forsyth, 2008)

Labeling because it 
 gives you added value Visipedia


(Belongie, Perona, et al)

Labeling for fun

Luis Von Ahn and Laura Dabbish 2004

Getting more humans in the annotation loop

Labeling to prove you’re human

33

Source: Isola, Torralba, Freeman

Beware of the human in your loop

• What do you know about them?
• Will they do the work you pay for?

Let’s check a few simple experiments

34

Source: Isola, Torralba, Freeman

People have biases…

Turkers were offered 1 cent to pick a number from 1 to 10.

From http://groups.csail.mit.edu/uid/deneme/ ~850 turkers Experiment by Greg Little

35

Source: Isola, Torralba, Freeman

Do humans have consistent biases?

Results form 100 HITS: From http://groups.csail.mit.edu/uid/deneme/ Experiment by Greg Little

36

Source: Isola, Torralba, Freeman

Are humans reliable even in simple tasks?

From http://groups.csail.mit.edu/uid/deneme/ Experiment by Greg Little

A: 2 B: 96 C: 2 Results of 100 HITS:

37

Source: Isola, Torralba, Freeman

Do humans do what you ask for?

31 heads, 19 tails After 50 HITS: 34 heads, 16 tails And 50 more:

From http://groups.csail.mit.edu/uid/deneme/ Experiment by Rob Miller

38

Source: Isola, Torralba, Freeman

But suppose we messed up. Our test setting doesn’t look like the training data! So we can sometimes collect good training data. How can we bridge the domain gap?

39

Source: Isola, Torralba, Freeman

Finding more representative images

[Fouhey et al., "From Lifestyle Vlogs to Everyday Actions", 2017]

Finding more representative images

[Fouhey et al., "From Lifestyle Vlogs to Everyday Actions", 2017]

42

Name that dataset game

[Torralba and Efros, “An unbiased look at dataset bias,” 2011]

Domain gap between and will cause us to fail to generalize.

ptrain

ptest

training domain testing domain

(where we actual use our model) Space of natural images Training data Test data

43

Source: Isola, Torralba, Freeman

Domain gap between and will cause us to fail to generalize.

source domain target domain

(where we actual use our model) Space of natural images Source data Target data

psource

ptarget

44

Source: Isola, Torralba, Freeman

Data space source data target data

Idea #1: transform the target domain to look like the source domain (Or vice versa) This is called domain adaptation

45

Source: Isola, Torralba, Freeman

Domain adaptation

• We have source domain pairs {xsource, ysource}
• Learn a mapping F: xsource —> ysource
• We want to apply F to target domain data xtarget
• Find transformation T: xtarget —> xsource
• Now apply F(T(xtarget)) to predict ytarget

46

Source: Isola, Torralba, Freeman

It’s a just another distribution mapping problem!

psource

ptarget

47

Source: Isola, Torralba, Freeman

Horses Zebras

X

Y

CycleGAN

48

Source: Isola, Torralba, Freeman

Domain adaptation

psource

ptarget

49

Source: Isola, Torralba, Freeman

Domain gap between and will cause us to fail to generalize.

Space of images

psource

ptarget

source domain

Source data

target domain

(where we actual use our model) Target data

50

Source: Isola, Torralba, Freeman

CyCADA: Cycle-Consistent Adversarial Domain Adaptation

Source domain

,

Target domain

, ?

[Hoffman, Tzeng, Park, Zhu, Isola, Saenko, Darrell, Efros, arXiv 2017]

51

Source: Isola, Torralba, Freeman

CycleGAN

,

Training data

52

Source: Isola, Torralba, Freeman

,

Training data

CycleGAN

53

Source: Isola, Torralba, Freeman

,

Training data

CycleGAN FCN

54

Source: Isola, Torralba, Freeman

OpenAI Dactyl

55

Source: Isola, Torralba, Freeman

Domain gap between and will cause us to fail to generalize.

Space of images

psource

ptarget

target domain

(where we actual use our model) Target data

source domain

Source data

56

Source: Isola, Torralba, Freeman

Data space Training data Test data

Idea #2: train on randomly perturbed data, so that test set just looks like another random perturbation This is called domain randomization or data augmentation

57

Source: Isola, Torralba, Freeman

Data augmentation

Training data

n

• ,

n

• ,

n

• ,

…

x

“Fish” “Grizzly” “Chameleon”

y

n

• “Fish”

Mirror

n

• “Fish”

Crop

n

• “Fish”

Crop

n

• “Fish”

Darken

58

Source: Isola, Torralba, Freeman

Domain randomization

Above example is from [Tobin et al. 2017]

Training data

Test data

[Sadeghi & Levine 2016]

59

Source: Isola, Torralba, Freeman

60

Source: Isola, Torralba, Freeman

What if we go way outside of the training distribution?

61

Source: Isola, Torralba, Freeman

Training data

X

Y

Test data

X

Y

{x(train)

i

, y(train)

i

}N

i=1

{x(test)

i

, y(test)

i

}M

i=1

Our training data did not cover the part of the distribution that was tested (biased data)

generalization error

62

Source: Isola, Torralba, Freeman

“Deep Neural Networks are Easily Fooled: High Confidence Predictions for Unrecognizable Images”

[Nguyen, Yosinski, and Clune, CVPR 2015]

63

Source: Isola, Torralba, Freeman

“Deep Neural Networks are Easily Fooled: High Confidence Predictions for Unrecognizable Images”

[Nguyen, Yosinski, and Clune, CVPR 2015]

64

Source: Isola, Torralba, Freeman

Data space Training data

Weirdness of high-dimensional space:

Usually, there are blind spots where the model has not fit the distribution well, and behaves unpredictably

65

Source: Isola, Torralba, Freeman

Adversarial noise

f

“Ostrich”

[“Intriguing properties of neural networks”, Szegedy et al. 2014]

x

+

r

=

x + r

f

“School bus”

y

arg max

r

p(y = ostrich|x + r) subject to krk < ✏

66

Source: Isola, Torralba, Freeman

Anything to worry about?

“NO Need to Worry about Adversarial Examples in Object Detection in Autonomous Vehicles”, Lu et al. 2017 (Early) 2017’s attacks fail on physical

• bjects, since they are optimized to

attack a single view!

67

Source: Isola, Torralba, Freeman

Anything to worry about?

3D-printed turtle model classified as rifle from most viewpoints “Synthesizing Robust Adversarial Examples”, Athalye, Engstrom, Ilyas, Kwok, 2017 Later in 2017…

68

Source: Isola, Torralba, Freeman

Adversarial examples

• Current deep models have bad worst-case performance
• Can be exploited by an adversary
• Few guarantees, can’t fully trust what the model’s output

69

Source: Isola, Torralba, Freeman

Mission-critical computer vision systems

Source: https://www.nytimes.com/2020/02/25/business/tesla-autopilot-ntsb.html, https://qz.com/1402843/watch-teslas-autopilot-see-the-streets-of-paris/

70

Some things to worry about…

• Misinformation
• Our datasets are often poorly labeled
• And usually biased (overrepresent certain categories)
• ML method perform beautifully on laboratory data, but often generalize

poorly to real-world data

• Can have negative social consequences

71

72

Project office hours until 2pm