Visual Thinking:
Unpacking Visuals in Machine Learning
Samir Passi
background
visual thinking in ML education
significance of visuals in understanding ML
visual thinking in ML corporate practice
two main challenges
questions
2
PhD candidate in information science
data science/ML Intern at IB
background
software engineering + STS + information science
research area
critical data studies
research
work and decisions that go into machine learning
3
focus
learning environment + corporations
methods
ethnography, interviews, and in-practice
goal
ways to teach ML more effectively, research human and
management problems of ML, and create better tools/strategies for
ML practices
short-term aim
through a description of visuals in ML,
showcase two key challenges in ML corporate practices
4
visual thinking
5
6
Visual thinking
(ML education)
7
8
9
10
11
all images are from the 12
internet (CC license)
approach
all images are from the 13
internet (CC license)
what is even more special about the visuals
used in machine learning and data science?
ways of
seeing
ML models
+
ways of
interacting
with ML models
16
17
18
19
20
21
learn
22
23
react
24
visuals help us to better understand
how our machine learning models:
approach problems,
learn rules,
react to evaluation metrics
25
visuals are indispensible for ML, right?
26
visual thinking
(ML corporate practice)
27
visual thinking
doesn’t scale very well
28
churn
prediction
1-feature
support page access
churn
prediction
1-feature
support page access
churn
prediction
1-feature 2-features
support page access
support page access
number of packages
churn
prediction
1-feature 2-features
support page access
support page access
number of packages
churn
prediction
1-feature 2-features 3-features
support page access
support page access
support page access
number of packages
number of packages
number of users
churn
prediction
1-feature 2-features 3-features
support page access
support page access
support page access
number of packages
number of packages
number of users
churn
prediction
1-feature 2-features 3-features 4-features? 50-features?
c
support page access
support page access
support page access
number of packages
number of packages
number of users
visual thinking
doesn’t scale very well
in corporate ML applications
36
problem #1
more data = better models
more data = visual challenges
better models = visual challenges (?)
37
problem #1
more data = better models
more data = visual challenges
better models = visual challenges (?)
38
problem #2
more data = better models
more data = visual challenges
better models = visual challenges (?)
visual challenges amplify explanation challenges
39
what can we do in the meantime?
ask different questions
40
‘why does the model say that this particular
customer is going to churn?’
‘what parameters does your model have, what do
they do, and how does your model make a decision?’
41
what can we do in the meantime?
explicitly articulate hard goals
42
‘we want to know which customers will churn’
‘maybe you want a highly specific churn-prediction
model, or perhaps cost-cutting is the main goal here,
or both?’
43
what can we do in the meantime?
voice soft goals a.k.a. concerns
44
‘we want to increase sales’
‘maybe you don’t want models to boost sales at the
expense of customer satisfaction, or perhaps you are
okay with a less accurate model that produces easy
explanations’
45
traditional and visual
forms of thinking
may not scale very well in corporate ML applications
46
traditional and visual
forms of thinking
may not scale very well in corporate ML applications
but, our question, goals, and concerns can…
47
thank you!
questions & comments?
This work has been funded by:
48
