The studio is a hallmark trait of design education and practice. Working in a shared space, students solicit each other’s help and gain wisdom by seeing their peers’ work and failures, successes, and evolutions. It’s a tremendously powerful learning experience. It’s also tremendously resource intensive. Studios generally require dedicated, collocated space for students, and studio classes tend to have extremely limited enrollment. How might we create online experiences that are inspired by the design studio, and open those peer learning opportunities to learners around the globe?
In this COIL Fischer Speaker Series event, Scott Klemmer shared his adventures in creating global-scale peer learning systems for formative feedback, small-group discussion, and summative assessment. In 2012, his research group collaborated with Coursera to launch the first massive-scale class with self and peer assessment. Since then, their systems have been used by more than a hundred massive online classes and on-campus flipped classrooms. Because online learning platforms embed pedagogy into software, they provide a powerful setting for using and building theory through experimentation.
Scott also used their online learning research as a case study in Design at Large: experiments and research systems leverage real-world, web-scale usage to create practical theories for design. Currently, many design practices are faith-based rather than research-based. Why is there a shortfall of principles? In part, some see design as intrinsically mystical and impervious to investigation, because creative work is clearly complex and multifarious. And in part, this is a multidisciplinary effort. Design is front-page news, the topic of Hollywood films, and enrollment in design courses — both in person and online — has skyrocketed. For Scott, the most powerful part is how many people are excited about making stuff. Let’s match this enthusiasm with insight.
The video of this presentation can be viewed at https://goo.gl/maJfh0
6. Norman & Klemmer (2014) How design education must change
Why a shortfall of design principles?
• Engineering excels at practical theory
…from the physical sciences.
• The human world is different
• Introspection is valuable
…but often misleading
• Industry is empirical
…but product focused
14. —E.W. Dijkstra, On the Cruelty of Really Teaching
Computer Science
Just for Small Innovations?
“By ... metaphors and analogies we try to link the new to the old,
the novel to the familiar. Under sufficiently slow and gradual
change, it works reasonably well;
in the case of a sharp discontinuity, however, the method breaks
down ... our past experience is no longer relevant, the analogies
become too shallow, and the metaphors become more
misleading than illuminating. This is the situation ... for radical
novelty.”
17. “Good artists borrow, great artists steal”
—Pablo Picasso
19th century Fang sculptureLes Demoiselles d'Avignon
John Richardson, A Life of Picasso:The Cubist Rebel, 1907-1916
19. Design Learning at Large
Chinmay Kulkarni et al.
Peer and Self Assessment in Massive Online Classes, Chinmay
Kulkarni, Koh Pang Wei, Huy Le, Daniel Chia, Kathryn Papadopoulos,
Justin Cheng, Daphne Koller, Scott R. Klemmer. TOCHI: ACM
Transactions on Computer-Human Interaction, 2013
The identify-verify pattern scales short-answer grading by combining
peer assessment with algorithmic scoring, Chinmay Kulkarni, Richard
Socher, Michael S. Bernstein, Scott R. Klemmer. Learning at Scale, 2014
Talkabout: Making distance matter with small groups in massive
classes, Chinmay Kulkarni, Julia Cambre, Yasmine Kotturi, Michael S.
Bernstein, Scott Klemmer, CSCW: ACM Conference on Computer
Supported Cooperative Work, 2015
Structure and messaging techniques for online peer learning systems
that increase stickiness, Yasmine Kotturi, Chinmay Kulkarni, Michael
Bernstein, Scott Klemmer, ACM Learning at Scale, 2015
PeerStudio: Rapid Peer Feedback Emphasizes Revision and Improves
Performance, Chinmay Kulkarni, Michael S Bernstein, Scott R Klemmer,
ACM Learning at Scale, 2015
23. 1. Feedback on open-ended work
Schön, D. (1987). Educating the reflective practitioner: Toward a new
design for teaching and learning in the professions.
24. 2. Engaging Diverse Perspectives
Model UN Design Crit
Gurin, P. et al. (2002) Diversity and higher education: Theory
and impact on educational outcomes, Harvard Educational Review
25. 3. Revision for mastery
Ericsson, K. A., Krampe, R. T., & Tesch-Römer, C. (1993). The role of
deliberate practice in the acquisition of expert performance.
Image Courtesy IDEO
26. Global peer assessment
Interfaces for accurate assessment of open-
ended work
Talkabout
Diversity as a design opportunity:
Small group discussions in massive classes
PeerStudio
Scale as a design opportunity:
Immediate feedback for mastery
27. The paradox of peer
processes
Non-experts performing expert work
28. Our approach:
Calibrated peer review
Chinmay Kulkarni, et al.
Peer and Self Assessment in Massive Online Classes, TOCHI 2013
3) Reflect
(Assess: Self)
2) Assess: Peers1) Train: calibrate
✓
30. Large scale peer assessment
Human-computer
Interaction
Design
Teaching
character
Management
Constitutional law
Arguments
Introduction to
Philosophy
Essays
Social
Psychology
Essays
Programming
in Python
Code
Child
Nutrition
Recipes
World Music
Music
used by 100,000+ students
31. Assessment training is crucial
0
25
50
75
100
0 25 50 75 100
Self grade (%)
Peergrade(%)
No Training
r=0.58
0
20
40
60
80
100
0 20 40 60 80 100
Self grade (%)
Peergrade(%)
With Training
r=0.73
32. How well do peer and staff
assessments correlate?
3) Reflect
(Assess: Self)
2) Assess: Peers
staff-graded
1) Assess: calibrate
✓
Dataset: 99 submissions with ~160 peer assessments each.
33. Grading for a pass-fail class
Extrapolated results from a bootstrapped simulation
Earn certificate if staff-graded
Certificate awarded
No certificate
97.8%
2.2%
No certificate
Certificate awarded
99.3%
0.7%
No certificate if staff-graded
34. Students with novel answers
sometimes penalized unfairly
“damn peer review - it was a bunch of
[students] just making things fit into a rubric -
checking off a check sheet - like talking about
dog poop. what is this world coming to?”
-A student in a peer-assessed class
35. “I've never seen something like that!”
Introduction to Art
“Treasure Cage” from Canada“Magical lights” from Norway
36. The return of the
novices-as-experts paradox
“fully interactive, page flow is
complete… make it clearer
what people should do next”
Experts:
capture the structure
of rubric
Peers:
Focus on superficial
features, even when
asked not to
“unpolished…Try to make UI less
coloured.”
37. Fortune cookies for qualitative,
personalized feedback
• Peers can recognize errors from a list of
patterns, even if they can’t articulate them
• Most errors are variations on a theme
+
“...because _____________________”
Cue Variation
38. Students Made it Theirs
• Sharing cool interfaces, resources,
articles
• Collating reading lists, creating
assignment aids
• Doing really creative work
• Helping other students
• heuristic evaluation feedback
• answering forum questions
• extra peer assessment
39.
40.
41.
42.
43. I am Chandramouli Sharma, a junior year undergraduate in Computer Science from the National Institute of Technology Karnataka, India.
I am one of those thousands of students who took the HCI class on Coursera in October 2012. I had timing clashes, so I had to finish the
course during December in vacations.
Here is my amazing journey from a small project in HCI class to a platform that will now be used by thousands of schools in 47 countries
and the awards I won along the way. I have illustrated it through pictures. This is a tribute to you for the great you class took at no fee.
Note: Images might take some time to load.
What I worked on..
I worked on a web application which could display complex environmental pollution data sets into interactive visualizations. This could be
used by school students to understand environmental issues. Below is one of the paper prototypes that I developed during the class.
After a few iterations I came up with a digital prototype. It looked something like this.
44. Global peer assessment
Interfaces for accurate assessment of open-
ended work
Talkabout
Diversity as a design opportunity:
Small group discussions in massive classes
PeerStudio
Scale as a design opportunity:
Immediate feedback for mastery
45. Why diversity?
Different professional
knowledge, educational
systems, and cultural
values
Information
[Tudge ’08]
Cognition
[Gurin et al. ’02]
[Nemeth ’86]
[Schwartz et al ’04]
From passivity to active,
effortful, conscious
thinking
46. Students are often homophilic
Hurtado, S. et al. (1998) The Climate for Diversity: Key Issues
for Institutional Self-Study.
47. Talkabout: video discussions
with global peers
Kulkarni, C, et. al. “Talkabout: Making distance matter with
small groups in massive classes”, CSCW 2015
48. Group assignment algorithm
• Talkabout assigns to one of many parallel
groups.
• Assignment is greedy, constrained by
preferred group size
• balances gender
• improves geographic diversity
50. Discussants as far apart as
New York and London
Median pair-wise distance 4,100 mi (6,600 km)
0
50
100
2 3 4 5 6
Number of countries in discussion
NumberofDiscussions
51. Students discuss twice as long
as instructors asked them to
Discussion duration (minutes)
Number of
Students
Median duration
0
100
200
30 60 90 120 150 180
Recommended
duration
52. Do diverse, small-group discussions
improve learning outcomes?
1. Does participation help?
2. Does diversity amplify
participation benefits?
IRB #30319
53. Study: Benefits of
Participation
• n=934, Irrational Behavior
• Dependent measure: total course grade (%)
• Between-subjects
Wait list
No talkabout for
first half of class
Discussion
Talkabout
throughout class
54. Course grades higher in
discussion condition
Irrational Behavior
(p<0.05)
Total
grade
(%)
0
10
20
30
40
50
Discussion Wait list
(control)
6% of total grade
55. Do diverse, small-group discussions
improve learning outcomes?
1. Does participation help?
2. Does diversity amplify
participation benefits?
56. Study: Benefits of Diversity
• n=2,422, Social Psychology
• Quasi-experiment: discussants assigned to
first available group
• Result: natural variation in diversity
• Measure: performance on final exam
• OLS regression controls for prior performance
57. Diverse discussions lead to
higher final scores
0%
2.5%
5%
Social Psychology Organizational Analysis
3.6%
2.4%
Grade
difference
(most-least
diverse)
ior
58. Evaluation Goals
Do diverse, small-group discussions
improve learning outcomes?
1. Does participation help?
2. Does diversity amplify
participation benefits?
60. Talkabout as a springboard to
global friendships
“We shared emails because we are discussing
issues that require a strong, networked group
to change the status quo… the impact would
be far greater if participants could connect
and engage outside of the course”
-Student in International Women’s Health and Human Rights class
Average (9 classes)
International Women’s
Health and Human Rights
0% 25% 50% 75% 100%
92%
47.2%
“Shared contact info with group”
61. 5,000+ students from 134 countries
Social Psychology
International Women’s
Health & Human
Rights
Learning How to Learn
How to Change the
World
Understanding
Research Methods
Irrational Behavior
Critical Perspectives
on Management
Organizational
Analysis
Think Again: How to
Reason and Argue
translated by students into French & Spanish
62. Global peer assessment
Interfaces for accurate assessment of open-
ended work
Talkabout
Diversity as a design opportunity:
Small group discussions in massive classes
PeerStudio
Scale as a design opportunity:
Immediate feedback for mastery
63. PeerStudio scales interactive
peer feedback
Kulkarni C., Bernstein M., Klemmer S. (2015)
“PeerStudio: Rapid Feedback Emphasizes Revision and Improves
Performance”, Learning@Scale
Submit for feedback
Give feedback to two
peers
Submit for
grades
Read feedback & revise
64. How might we lower the
training burden?
0.0
2.5
5.0
7.5
10.0
Evaluation Submission
10.5 hours
1.9 hours
Training Creating
own work
Median hours
in activity
Training
1.9 hours
65. Solution: contrasting cases for
training-free micro-expertise
Thompson, Gentner, Loewenstein (2000),
“Analogical Training More Powerful Than Individual Case Training”
Average Peer-majority/Staff
difference: 5.7%
66. Time to first feedback:
Learning How to Learn
0
50
100
<10 min <1 hr <2 hr <6 hr <12 hr <24 hr > 24 hr
Time to first review
Numberofsubmissions
native the plot
68. Solution: Real-time tips for
actionable feedback
• Correctness and velocity feedback leads to
large improvements
• Specific, topic-relevant feedback more useful
• Logistic regression with bag-of-words features
predicts relevance
69. Solution: Real-time tips for
actionable feedback
1 Calculate an internal score
for each rubric dimension
2 Generate tips for reviewer
Overall, 81% of students
received actionable
comments
70. Without hints, students focus on
author, and what’s good
I think you are, I wish you, I hope you…
With hints, students focus on
work and what could be better
I think you should, you need to,
your work could…
71. N=104 in “Medical Education in the New Millennium” (edX)
Study: Does fast feedback
improve final performance?
Early feedback,
fast (<1 hr)
grades 4.4%
higher than No early
feedback
Early feedback,
delayed 24 hours
No early
feedback
grades same
as
72. Global peer assessment
Interfaces for accurate assessment of open-
ended work
Talkabout
Diversity as a design opportunity:
Small group discussions in massive classes
PeerStudio
Scale as a design opportunity:
Immediate feedback for mastery
73. • Build practical theory with
real-world experiments
• Bake pedagogy into
software that transforms
learning
SCALING THE STUDIO
http://d.ucsd.edu/peer
81. Thomke (2000) Experimentation matters: unlocking the potential of new technologies for innovation
Learning through Prototyping
“Never go to a meeting
without a prototype...”
—Boyle’s Law
82. Design Process at Large
Steven Dow
Asst Prof, CMU
Early and Repeated Exposure to Examples Improves Creative Work,
Chinmay Kulkarni, Steven P Dow, Scott R Klemmer. Cognitive
Science, 2012.
Prototyping Dynamics: Sharing Multiple Designs Improves
Exploration, Group Rapport, and Results, Steven P Dow, Julie
Fortuna, Dan Schwartz, Beth Altringer, Daniel L Schwartz, Scott R
Klemmer. CHI: ACM Conference on Human Factors in Computing
Systems, 2011.
Parallel Prototyping Leads to Better Design Results, More
Divergence, and Increased Self-Efficacy, Steven P Dow, Alana
Glassco, Jonathan Kass, Melissa Schwarz, Daniel Schwartz, Scott R
Klemmer. ACM Transactions on Computer-Human Interaction, 2010
The Efficacy of Prototyping Under Time Constraints, Steven P. Dow,
Kate Heddleston, Scott R Klemmer. Creativity & Cognition, 2009
83.
84. “I went with the whole parachute idea and what I had from the
beginning...”
“This is the best approach for such a design...” “I am not a very good outside-the-box thinker, so I kinda just had one idea
and I was going to try to make it work...”
“No... for some reason... this seems to be the only idea. There needs to be a
platform and then as good of cushion as possible... I don’t see any other way.”
Participants picked their concept early
88. Task: Design a Web Ad (N=33)
parallel
prototyping
condition
FINAL
serial
prototyping
condition
89.
90. Parallel design -> more clicks
Parallel
Clicks per million
impressions
Serial F(1,30)=4.227
p<.05
0
60
120
180
240
300
360
420
480
398
445
91. ...and more time on the site
Parallel
condition
Average time on client
site per visitor
(seconds)
Serial
condition
F(1,493)=3.172
p=0.076
0
5
10
15
20
25
30
35
40
12.9
31.3
93. ...and more diverse designs
Parallel Serial
F=182, p<0.001
0
0.5
1
1.5
2
2.5
3
3.5
3.18
2.78
7=highly similar
0=not at all similar
94. Gentner, Loewenstein, & Thomson, 2003
learning outcome
Comparison aids learning
training
session
“Describe the solution.”
CASE#1
CASE#2
CASE#1
CASE#2
“Describe the parallels of
these solutions”
“Describe the solution.”
SEPARATE CASES COMPARISON CASES
Solutions to a landlord-renter lease
~ 3x
95. Sharing Multiple Benefits
• User engagement
• Expert rating
• Individual exploration
• Feature sharing
• Conversational turns
• Consensus
• Rapport Share Multiple
clicks/M
0
250
500
750
1000
1250
774.6734.9
1072.1
Share Best Share One
χ2=4.72, p<0.05
self other self other
… …
self other