Ethics for Data Scientists
The Limits of ML
Munich DataGeeks
Gerrit Gruben
31. January 2018
Slide 2
Slide 2 text
about.me
2
•Freelance DS, before worked as
DS/SWE.
•Training people in a 3-month
boot camp to be DS →
•Org. of Kaggle Berlin meetup
•ML PhD Dropout @ Potsdam
•Degrees in Math. & CS, going for
Laws (sic!)
datascienceretreat.com
Slide 3
Slide 3 text
Goals
3
Slide 4
Slide 4 text
4
Slide 5
Slide 5 text
Main points
•No data positivism in ML
• inductive bias always there
• IID assumption is idealistic.
•Can't predict everything
•ML systems prone to manipulation (fragility)
5
Slide 6
Slide 6 text
Limits & Biases
6
Slide 7
Slide 7 text
7
Benevolent or evil?
Slide 8
Slide 8 text
”Absence of Evidence is not Evidence of
Absence" --- Data Scientist’s Proverbs
”I beseech you, in the bowels of Christ, think it
possible that you may be mistaken" --- Oliver
Cromwell
Dennis Lindley: avoid prior probabilities of 0
and 1.
Slide 13
Slide 13 text
Problem of Induction
•More general as the black swan problem.
•ML models have an inductive bias.
13
” The process of inferring a general law or
principle from the observation of particular
instances." --- Oxford's Dictionary
(direct opposite of deduction)
Slide 14
Slide 14 text
” When you have two competing theories that
make exactly the same predictions, the simpler
one is the better." --- Ockham’s Razor
Slide 15
Slide 15 text
Technical Things
What goes wrong often…
Slide 16
Slide 16 text
Multiple Testing
Retrying the tests so often, until "hitting" the significance
level by chance.
Solution: Bayesian or correction (e.g. Bonferroni correction)
or different experimental design.
Data Snooping: http://bit.ly/2iWoFrV
"P-hacking" II
"When a measure becomes a target, it ceases to
be a good measure" --- Goodhart's law
Slide 22
Slide 22 text
selection
≠ evaluation
Slide 23
Slide 23 text
23
Paper: http://bit.ly/2gBIR1M
Prefer to call it
“over-selection”
In “Learning with Kernels”
from Smola & Schölkopf they
name ex. 5.10. “overfitting on
the test set”.
Slide 24
Slide 24 text
Empirical Risk Minimization
•
24
Slide 25
Slide 25 text
Empirical Loss
•
25
Slide 26
Slide 26 text
Empirical Risk Minimization II
•
26
Slide 27
Slide 27 text
Bias / Variance
27
Slide 28
Slide 28 text
•
28
Slide 29
Slide 29 text
29
Source: http://bit.ly/2vDfoLp
Slide 30
Slide 30 text
30
Source: University of Potsdam
Slide 31
Slide 31 text
31
Source: University of Potsdam
Slide 32
Slide 32 text
Nested CV
32
From Quora: http://bit.ly/2wvz2aZ
Slide 33
Slide 33 text
Messing up your experiments
•Data split strategy is part of experiment.
•Mainly care for:
• Class distribution
• Problem domain relevant issues such as time
33
”Validation and Test sets should model nature
and nature is not accommodating." --- Data
Scientist’s Proverbs
Slide 34
Slide 34 text
34
“Model evaluation, model selection…“
by Sebastian Raschka: http://bit.ly/2p6PGY0
“Approximate Statistical Tests For
Comparing Supervised Class. Learning
Algorithms” (Dietterich 98):
http://bit.ly/2wyItF6
Slide 35
Slide 35 text
Gallery of Fails
Slide 36
Slide 36 text
Courier/Terrorist detection in Pakistan
36
Source: http://bit.ly/1KY4SQE
Slide 37
Slide 37 text
Feedback loops abused
Tay.ai was a chat bot
deployed on Twitter
by Microsoft for just a
day.
Trolls started to
"subvert" the bot by
"teaching" it to be
politically incorrect by
focussed exposure to
extreme content.
Slide 38
Slide 38 text
Moral Machine http://moralmachine.mit.edu
38
Slide 39
Slide 39 text
Smaller tips for ML
•Always model uncertainty.
•Read this
•Don’t mock values of a
non-existant predictive model.
39
Slide 40
Slide 40 text
Books
Slide 41
Slide 41 text
Other Links
•https://www.ma.utexas.edu/users/mks/statmistakes/StatisticsMist
akes.html
•Quantopian Lecture Series: p-Hacking and Multiple Comparison
bias https://www.youtube.com/watch?v=YiDfbYtgUPc
•David Hume: A Treatise on Human Nature:
http://www.davidhume.org/texts/thn.html
41