1 of 85

On Bullshit,

and beyond

Giovanni Colavizza

2 of 85

Announcements

Please fill in your group info to prepare the poster session (more info soon): https://docs.google.com/spreadsheets/d/11mFYtoONgLYR4Dky3CERYi4rRBVox-TGbB9MVu-YJ1A/edit?usp=sharing

Tomorrow:

  • Talk by Pierre Dillenbourg (EPFL CHILI)
  • Project office hours (last before deadline!)

Tue 19th 23:59 CET hard deadline for Project M3

Next (last) week:

  • Class on Networks
  • Lab with talk on networks and primer on licenses

3 of 85

“One of the most salient features of our culture is that there is so much bullshit. Everyone knows this. Each of us contributes his share. But we tend to take the situation for granted.”

Frankfurt, On Bullshit, 2005 [1986] https://www.stoa.org.uk/topics/bullshit/pdf/on-bullshit.pdf

4 of 85

Fake news

5 of 85

Fake news

6 of 85

Bias and discrimination

7 of 85

Sensationalization of scientific results

8 of 85

Rise of terminator machines and the like

9 of 85

Everyday’s bullshit

10 of 85

This class

Essentially about self-awareness

Two parts:

  • Tactics to counteract bullshit in the digital world
  • The work of a data scientist in context

Cf. "Calling Bullshit"

http://callingbullshit.org/index.html

11 of 85

What is bullshit, anyway?

12 of 85

What is bullshit, anyway?

Bullshit as a fundamental disregard for truth

(liar != bullshitter)

Frankfurt, On Bullshit, 2005 [1986] https://www.stoa.org.uk/topics/bullshit/pdf/on-bullshit.pdf

13 of 85

What is bullshit, anyway?

Bullshit as a fundamental disregard for truth

(liar != bullshitter)

  • Bullshitter acts as if s/he is more certain than warranted

Bullshit as a mental state.

Meibauer, Aspects of a Theory on Bullshit, 2016

14 of 85

What is bullshit, anyway?

Bullshit also about unverifiable unclarity

Bullshit as an ontological entity.

Cohen, Deeper Into Bullshit, 2002

http://learning.hccs.edu/faculty/robert.tierney/phil1301-6/bullshit/g.a.-cohen-deeper-into-bullshit/view

15 of 85

What is bullshit, anyway?

16 of 85

Cargo Cult Science

17 of 85

Bullshit goes digital

18 of 85

Brandolini’s asymmetry principle

“The amount of effort necessary to refute bullshit is one order of magnitude bigger than to produce it.”

19 of 85

Is bullshit getting worst?

20 of 85

Is bullshit getting worst?

21 of 85

Is bullshit getting worst?

22 of 85

Tactics for fighting bullshit

23 of 85

3 parts

  1. Spotting bullshit
  2. Statistical pitfalls
  3. Data visualization

24 of 85

Spotting bullshit

  • Don’t fall in love with your hypothesis
  • Occam’s razor
  • Who, how, what’s in it for them…
  • Check the sources
  • Too good to be true
  • Confirmation bias
  • Multiple working hypotheses
  • Orders of magnitude
  • Unfair comparison

25 of 85

Fox: 70M$ wasted in stamp fraud last year (2016)

26 of 85

Orders of magnitude: Fermi estimation

Fraud: 70M$

Prop. of Americans on food stamps: ˜10%

Dollars per american/year: ˜1000

Total program: ˜30'000M

Fraud: ˜0.2%!!

27 of 85

More...

How to spot fake news (FactCheck): http://www.factcheck.org/2016/11/how-to-spot-fake-news/

Tim O’Reilly, How I detect fake news: https://www.oreilly.com/ideas/how-i-detect-fake-news

28 of 85

Statistical pitfalls

  • Causation and correlation
  • Prosecutor’s fallacy
  • Right censoring
  • Averages and medians
  • Garbage-in, garbage-out
  • Simpson’s paradox
  • Biased sampling
  • Sensible trends
  • Overfitting
  • Big data hubris

29 of 85

Causation and correlation

30 of 85

Prosecutor's fallacy

31 of 85

Prosecutor's fallacy

Match

No match

Guilty

1

0

Innocent

8

7'999'992

P(match given innocent) = 8/8'000'000

P(guilty given match) = 1/9!!

32 of 85

Prosecutor's fallacy

Reject Ho

Don't reject

Ho false

True positive

False negative (Type II)

Ho true

False positive (Type I)

True negative

P(FP/(TN+FP)) = p-value!

P(TP/(TP+FP)) = depends on the alternative H we want to test (this is the statistical power)

P(H1 | reject Ho): even with very low p-value and high power, the alternative hypothesis (i.e. guilty) could be quite unlikely!

33 of 85

Right censoring

34 of 85

Garbage-in,

garbage-out

35 of 85

Simpson's paradox

36 of 85

37 of 85

Biased sampling

(+ cohort effects)

38 of 85

39 of 85

Sensible trends

40 of 85

Sensible trends

41 of 85

Data visualization done wrong

  • The y-axis
  • Binning
  • Unreadability
  • Missing proportionality
  • Improper scaling
  • Aberrations

Also cf. http://callingbullshit.org/tools.html

42 of 85

43 of 85

The y-axis

44 of 85

The y-axis

45 of 85

Wall Street Journal, April 17th 2011.

46 of 85

Binning

47 of 85

Unreadability

48 of 85

Missing proportionality

49 of 85

Improper scaling

50 of 85

Aberrations

51 of 85

Aberrations

52 of 85

Tufte's rules

"It is right to decorate construction, but never to construct decoration"

http://www.sealthreinhold.com/school/tuftes-rules/rule_one.php

53 of 85

Sagan’s toolkit

  • Start with a skeptical attitude
  • Check and seek independent confirmation of the facts
  • Encourage debate with knowledgeable parties
  • Do not put too much premium on arguments from authority
  • Spin multiple hypotheses
  • Do not get too attached to an hypothesis (yours)
  • Quantify and experiment
  • Seek to confirm all links in a chain of arguments
  • Apply Occam’s razor
  • Prefer falsifiable hypotheses

Sagan, The Fine Art of Baloney Detection,

http://www.inf.fu-berlin.de/lehre/pmo/eng/Sagan-Baloney.pdf

54 of 85

Sagan’s fallacies

  • Arguments ad hominem
  • Arguments from authority
  • Arguments from adverse consequences
  • Appeal to ignorance (whatever has not been proven false, must be true)
  • Assuming the answer
  • Observational selection (F. Bacon: counting the hits and forgetting the misses)
  • Statistics from small numbers
  • Inconsistency
  • Non sequitur
  • Post hoc, ergo propter hoc (temporal sequentiality implying causation)�

Sagan, The Fine Art of Baloney Detection,

http://www.inf.fu-berlin.de/lehre/pmo/eng/Sagan-Baloney.pdf

55 of 85

Cognitive biases

56 of 85

In the end, why is it a problem?

  • Not a new phenomenon, but likely at a new scale
  • Take it as information pollution
  • Great equalizer
  • Age of attention

57 of 85

The work of a data scientist, in context

58 of 85

Data science in the wild

Data science and ML work

(data crunching, analysis, models, tools, …)

Infrastructure

WORLD

data

WORLD

products, insights, policies,

...

59 of 85

Data science in the wild

Infrastructure

WORLD

data

WORLD

products, insights, policies,

...

60 of 85

Data science in the wild

Data science and ML work

(data crunching, analysis, models, tools, …)

Infrastructure

WORLD

data

WORLD

Products, insights, policies,

...

61 of 85

Let’s take a different perspective

analysis

WORLD

data

WORLD

action

62 of 85

3 stories

  1. Data
  2. Analysis
  3. Action

63 of 85

Cesare Lombroso’s positive criminology (data)

64 of 85

65 of 85

“We study, for the first time, automated inference on criminality based solely on still face images, which is free of any biases of subjective judgments of human observers.”

https://arxiv.org/pdf/1611.04135.pdf

66 of 85

Convicted persons’ IDs

Photos crawled from the Web

67 of 85

68 of 85

A smile will save you!

Black Mirror S03E01 “Nosedive”

69 of 85

Hubris is in the air…

“Unlike a human examiner/judge, a computer vision algorithm or classifier has absolutely no subjective baggages [sic], having no emotions, no biases whatsoever due to past experience, race, religion, political doctrine, gender, age, etc., no mental fatigue, no preconditioning of a bad sleep or meal. The automated inference on criminality eliminates the variable of meta-accuracy (the competence of the human judge/examiner) all together.”

70 of 85

Big

picture

If interested, check Kate Crawford’s talk at NIPS 2017: https://www.facebook.com/nipsfoundation/videos/1553500344741199/

AI Now Institute: https://ainowinstitute.org/

71 of 85

Big

picture

72 of 85

Google Flu Trends (analysis)

Detecting influenza epidemics using search engine query data,

doi:10.1038/nature07634

73 of 85

Google Flu Trends (analysis)

The Parable of Google Flu: Traps in Big Data Analysis, http://science.sciencemag.org/content/343/6176/1203

74 of 85

Google Flu Trends (analysis)

The Parable of Google Flu: Traps in Big Data Analysis, http://science.sciencemag.org/content/343/6176/1203

  • Overfitting (50M queries on thousands of observations)
  • External influences (media flu coverage, changes in Google search)
  • Lack of performance wrt other simpler models (time series regression)

75 of 85

University rankings (acting)

Launched by US News in 1983.

76 of 85

University rankings (acting)

How to they work?

  • Interviews
  • Feature engineering

Main transition over the years: from input to outputs

What could possibly go wrong?

77 of 85

Feedback loops

“U.S. News’s first data-driven ranking came out in 1988, and the results seemed sensible. However, as the ranking grew into a national standard, a vicious feedback loop materialized. The trouble was that the rankings were self-reinforcing.

“When you create a model from proxies, it is far simpler for people to game it. This is because proxies are easier to manipulate than the complicated reality they represent.”

Cathy O’Neil, Weapons of Math Destruction, 2016.

78 of 85

What you don’t measure...

79 of 85

Measurements in the wild

When a measure becomes a target, it ceases to be a good measure.

Goodhart’s law

The more any quantitative social indicator is used for social decision-making, the more subject it will be to corruption pressures and the more apt it will be to distort and corrupt the social processes it is intended to monitor.

Campbell’s law

80 of 85

By the way

81 of 85

Interconnectedness and cascading effects

“There can be errors in these systems which propagate very quickly. Because of the scale of their action space--they can be hitting a billion or two billion users per day--that means the costs of getting it wrong are very very high.”

Mustafa Suleyman - DeepMind

82 of 85

Historical social networks detour

“Bring the world closer together”

83 of 85

Conclusions

Data science and AI are great! But…

Keep questioning

Be aware

Develop an ethical stance

84 of 85

A final message

“The first principle is that you must not fool yourself, and you’re the easiest person to fool.”

Richard Feynman

85 of 85

Credits

A lot is taken from the University of Washington class “Calling Bullshit” (Carl T. Bergstrom and Javin West): http://callingbullshit.org/index.html