Article: http://bit.do/anttipaper
COMPUTATIONAL DESIGN OF
user interfaces
Antti Oulasvirta
Associate Professor
userinterfaces.aalto.fi
COMPUTATIONAL DESIGN
COMBINES DESIGN THINKING WITH
COMPUTATIONAL SCIENCES
(AND BEHAVIORAL/COGNITIVE SCIENCES)
COMPUTATIONAL APPROACHES TO UID
OVERVIEW
Design generation
Design
refinement
TODAY
Choose a design
δ1
Algorithm
…
Interaction
UI (δ)
δn
User
Design
adaptation
UI design
2.
1. Gallery of results
Model
(θ)
Model-based
combinatorial
Observations
3. What next?
Theories; Prior data; Opinion
“ZOO”
SOME STATE-OF-THE-ART RESULTS IN COMPUTATIONAL DESIGN
[Proc DIS’14]
PianoText: Optimizing Skill Transfer
CD GAIN ADAPTATION USING INVERSE
MODELING AND LOCAL OPTIMIZATION
ATTENTION-OPTIMIZED SKIM READING
WITH HEURISTIC ADAPTATION
Title
[Proc CHI’16]
Keyboard design as the assignment problem
Qwerty:
28 wpm
KALQ
37 wpm
< Frequent side changes >
[Oulasvirta et al. CHI’13]
Gesture set design using model-based
black box optimization
[Sridhar et al. CHI’15]
9
LAYOUT REFINEMENT USING AN
ENERGY MINIMIZATION APPROACH
DesignScape
U Toronto
ERROR REDUCTION USING STATE MACHINES
GAME LEVEL DESIGN
USING HUMAN PERFORMANCE MODELS
EASY
MEDIUM
DIFFICULT
Real gaming performance by a colleague
[Proc CHI’16]
SCATTERPLOT ENHANCEMENT USING PERCEPTUAL METRICS
Matplotlib default design
[Micallef et al. IEEE TVCG 2017]
MENU SYSTEM GENERATION USING MODEL-BASED
COMBINATORIAL OPTIMIZATION (ANT COLONIES)
Body Level One
[Proc. UIST’13]
APPLICATION MENU OPTIMIZATION
CASE: WINDOWS PHONE
Baseline
3000
2500
2000
Baseline
Evaluation
Optimised
Baseline
Optimised
3500
3
Optimised
Baseline
Users (N=16) were
significantly faster
3000
with the optimized
3
menu. While they
rated it to be more
2500
cluttered, they liked
the colors more.
2000
All
All except first
Mean selection time (ms)
Optimised
Mean selection time (ms)
3500
Mean selection time (ms)
Baseline
2
2
All
All except first
[Todi et al. DIS’16]
WEB PAGE DESIGN
USING MULTI-OBJECTIVE COMBINATORIAL OPTIMIZATION
Heading goes here
Button
Button
Button
www.firstlink.com
www.secondlink.io
www.thirdlink.org
[Proc DIS’16]
DATA-DRIVEN “DESIGN MINING” USING
STATISTICAL FEATURE MODELS
Ranjitha Kumar et al. (2013): a
pattern miner that suggests user
interfaces to queries.
top nav
topbar
nav bar
largelarge
header
header
largelarge
text text
blockblock
layout
layout
query
query
FigureFigure
13. Five
13.ofFive
the 20
of the
search
20 search
resultsresults
for the for
three-part
the three-part
DQL layout
DQL layout
query visualized
query visualized
on the on
left.theThe
left.query,
The query,
which executed
which executed
in 2.1 minutes,
in 2.1 minutes,
returnsreturns
pages pages
that share
that ashare
common
a common
high-level
high-level
layout, layout,
but exhibit
but exhibit
different
different
designs.
designs.
comparison,
comparison,
FigureFigure
15 shows
15 shows
nearest-neighbor
nearest-neighbor
resultsresults
for thefor the
top query
in Figure
in Figure
14 using
14 using
only the
only
vision-based
the vision-based
GIST GIST
de- deWebzeitgeist
Webzeitgeist
also enables
also enables
a newa kind
new of
kind
design-based
of design-based
ma- ma- top query
scriptors.
WhileWhile
these these
elements
elements
are visually
are visually
reminiscent
reminiscent
of
of
chine chine
learning.
learning.
For the
Forfirst
thetime,
first time,
applications
applications
can stream
can stream scriptors.
the query,
they bear
theylittle
bearstructural
little structural
or semantic
or semantic
relation
relation
to it. to it.
structured
structured
visualvisual
descriptors
descriptors
for page
for elements
page elements
from afrom
cen-a cen- the query,
Machine
Machine
Learning
Learning
POSTER GENERATION USING
PROBABILISTIC INFERENCE
Image Parsing via Stochastic Scene Grammar
ng
y Hierarchical
scene
s (clusters) from
al relations.
)
Image Parsing via Stochastic Scene Grammar
Yibiao Zhao, Song-Chun Zhu
1
Y
Department of Statistics University of California, Los Angeles Los Angeles, CA 90095, Department of
Statistics and Computer Science University of California, Los Angeles Los Angeles, CA 90095
Yibiao Zhao and Song-Chun Zhu
University of California, Los Angeles
Introduction
Stochastic Scene Grammar
This paper proposes a parsing algorithm for indoor scene
understanding which includes four aspects: computing 3D
scene layout, detecting 3D objects (e.g. furniture), detecting
2D faces (windows, doors etc.), and segmenting the
background. The algorithm parse an image into a hierarchical
structure, namely a parse tree. With the parse tree, we
reconstruct the original image by the appearance of line
segments, and we further recover the 3D scene by the
geometry of 3D background and foreground objects.
The grammar represents compositional structures of visual entities, which includes three types
of production rules and two types of contextual relations:
• Production rules: (i) AND rules represent the decomposition of an entity into sub-parts; (ii)
SET rules represent an ensemble of visual entities; (iii) OR rules represent the switching
among sub-types of an entity.
• Contextual relations: (a) Cooperative + relations represent positive links between binding
entities, such as hinged faces of a object or aligned boxes; (b) Competitive - relations
represents negative links between competing entities, such as mutually exclusive boxes.
P T (I(Av ))
v2Cl T
9
Introduction
This paper proposes a parsing algorithm for indoor scene understanding which
includes four aspects: computing 3D scene layout, detecting 3D objects (e.g.
furniture), detecting 2D faces (windows, doors etc.), and segmenting the
background. The algorithm parse an image into a hierarchical structure, namely a
parse tree. With the parse tree, we reconstruct the original image by the
appearance of line segments, and we further recover the 3D scene by the
geometry of 3D background and foreground objects.
Stochastic Scene Grammar
The grammar represents compositional structures of visual entities, which
includes three types of production rules and two types of contextual relations:
o Production rules: (i) AND rules represent the decomposition of an entity
into sub-parts; (ii) SET rules represent an ensemble of visual entities; (iii) OR
rules represent the switching among sub-types of an entity.
o Contextual relations:
between binding entities, such as hinged faces of a object or aligned boxes; (b)
such as mutually exclusive boxes.
(3)
usters) in each
guration
)|AG (v)).
(4)
[1], Hedau et
n the UIUC
Bayesian Formulation
We define a posterior distribution for a solution (a parse tree) pt conditioned on an image I.
This distribution is specified in terms of the statistics defined over the derivation of production
rules.
Y
Y
P(pt|I) / P(pt)P(I|pt) = P(S)
P(Chv |v)
P(I|v)
(1)
v2V n
v2V T
Bayesian Formulation
The probability is defined on the Gibbs distribution: and the energy term is decomposed as
three potentials:
E(pt|I) =
3D synthesis of novel views based on the parse tree
X
E OR (Ar(Chv )) +
v2V OR
X
E AND (AG (Chv )) +
v2V A ND
X
E T (I(⇤v ))
We define a posterior distribution for a solution (a parse tree) pt conditioned on
an image I. This distribution is specified in terms of the statistics defined over the
derivation of production rules.
(2)
⇤v 2⇤I ,v2V T
The probability is defined on the Gibbs distribution:
and the energy term is decomposed as three potentials:
Inference by Hierarchical Cluster Sampling
3D synthesis of novel views based on the parse tree
We design an efficient MCMC inference algorithm, namely Hierarchical cluster sampling, to
search in the large solution space of scene configurations. The algorithm has two stages:
Results
Results
• Clustering: It forms all possible higher-level structures (clusters) from lower-level entities by
production rules and contextual relations.
P+ (Cl|I) =
Y
v2Cl OR
P OR (Ar(v))
Y
AND
P+
(AG (u), AG (v))
u,v2Cl AND
Y
P T (I(Av ))
Inference by Hierarchical Cluster Sampling
We design an efficient MCMC inference algorithm, namely Hierarchical cluster
sampling, to search in the large solution space of scene configurations. The
algorithm has two stages:
o Clustering: It forms all possible higher-level structures (clusters) from lowerlevel entities by production rules and contextual relations.
(3)
v2Cl T
• Sampling: It jumps between alternative structures (clusters) in each layer of the hierarchy to
find the most probable configuration (represented by a parse tree).
Q(pt ⇤ |pt, I) = P+ (Cl ⇤ |I)
Y
AND
P(AG (u)|AG (v)).
(4)
u2Cl AND ,v2pt AND
o Sampling: It jumps between alternative structures (clusters) in each layer of
the hierarchy to find the most probable configuration (represented by a parse
tree).
Experiment and Conclusion
Segmentation precision compared with Hoiem et al. 2007 [1], Hedau et al. 2009 [2], Wang et
al. 2010 [3] and Lee et al. 2010 [4] in the UIUC dataset [2].
Experiment and Conclusion
Segmentation precision compared with Hoiem et al. 2007 [1], Hedau et al. 2009
[2], Wang et al. 2010 [3] and Lee et al. 2010 [4] in the UIUC dataset [2].
e hierarchical
m fast inference
xtual relations.
Compared with other algorithms, our contributions are
o A Stochastic Scene Grammar (SSG) to represent the hierarchical structure of
visual entities;
o A Hierarchical Cluster Sampling algorithm to perform fast inference in the SSG
model;
o Richer structures obtained by exploring richer contextual relations.
Compared with other algorithms, our contributions are
• A Stochastic Scene Grammar (SSG) to represent the hierarchical structure of visual
entities;
• A Hierarchical Cluster Sampling algorithm to perform fast inference in the SSG model;
• Richer structures obtained by exploring richer contextual relations. Website:
http://www.stat.ucla.edu/ ybzhao/research/sceneparsing
Website: http://www.stat.ucla.edu/~ybzhao/research/sceneparsing
TEMPLATE DESIGN © 2008
www.PosterPresentations.com
(b) Our result
(c) Original poster[35]
[Qiang et al. arXiv 2017]
COMPUTER-ASSISTED UI SKETCHING USING MODELBASED INTERACTIVE COMBINATORIAL OPTIMIZATION
[Proc. DIS’16]
Choice architecture adaptation using MPI and ML
Functionality selection using MIP
Input modality and interface design using RL
optimization
COMPUTATIONAL INTERFACE DESIGN
A CONFLUENCE OF METHODS
•
Combinatorial optimization - see http://bit.do/anttipaper
•
Bayesian inference
•
Reinforcement learning
•
Simulations
•
Control theory
•
Process algebra
•
State machines
•
Petri nets
A WINDOW OF OPPORTUNITY
•
•
Large potential gains: Internationally: Interaction design core to every major company. Finland: strong interaction and graphic design profession. USA:
Increasing investments into “design intelligence”
Maturing methodology: Significant improvements:
•
More and more examples and design problems tackled
•
First evidence for equal- or supra-human-level design
•
Enablers: Hardware and software is there
•
HELSINKI: Strong in computational sciences and design!
BEWARE!
•
Misguided hype weils AI and design
•
First disappointments are in (e.g., grid.io)
•
Harder than it appears!
•
Realistic problems complex to formulate
•
Issues with data and models of users
MODEL-BASED COMBINATORIAL OPTIMIZATION
IN 5 MINUTES
OUTLOOK
COVER FEATURE OUTLOOK
Contacts
Settings
Phone
Messages
Mail
Browser
App store
Maps
News
(b)
Alarm
Skype
Photoshop Stopwatch
Recorder
Skype
Dropbox
YouTube
(a)
User Interface Design
with Combinatorial
Optimization
Facebook
(c)
(d)
Antti Oulasvirta, Aalto University
of engineering design, so why not user interface design? The
author reviews progress and challenges in model-driven
UI optimization, in which an optimizer utilizes predictive
models of human perception, behavior, and experience to
anticipate users’ responses to computer-generated designs.
O
40
COM PUTE R PUBLISHED BY THE IEEE COMPUTER SOCIET Y
(f)
FIGURE 1. Examples of designs generated with recent model-based approaches to user interface (UI) optimization: (a) application
menu design, where the positions, colors, and sizes of icons are optimized for visual search, motor performance, color harmony, and
minimum clutter; (b) menu system design, where a hierarchy of items is organized into a menu optimized for visual search time and
consistency with other menus; (c) gestural input design, where commands are mapped to hand gestures to optimize for performance,
individuation of fingers, and learnability; (d) widget GUI design, where widget groups, types, and positions are decided based on users’
motor capabilities; (e) scatterplot design, where aspect ratio, marker sizes, and transparency are optimized for maximum performance
in tasks like correlation estimation by using perceptual models of brightness, contrast, and unit perception; and (f) web layout design,
where the elements of a page are positioned, colored, and sized to optimize visual search time, motor performance, color harmony,
and minimum clutter.
Optimization methods have revolutionized almost every field
ptimization methods in user interface (UI)
design is a long-standing topic in human–
computer interaction (HCI) research. Currently, user-centered design is largely
focused on human creativity, sensemaking, empathy,
and creation of meaning,1 but optimization methods have been explored as supplemental ways to help
speed up the design cycle and improve design quality
(for more information about computational UI design,
see the sidebar). Unlike any other design method in
HCI, optimization methods offer a greater-than-zero
chance of finding an optimal design, and some exact
methods even offer mathematical guarantees for solutions.2 More importantly, if the design task and its
(e)
assumptions are appropriate, the outcome represents
the best achievable design.
Optimization methods do not stop at searching for a
single best design; they can also be used to explore the
design space for alternative ideas. In addition, optimization could promote a change in design culture by encouraging the explication, scrutiny, and accumulation of
design knowledge that has tended to be tacit. However,
progress has been slow. Scarcely 15 years ago, keyboard
layout was the prime—and virtually only—application
of optimization methods that had empirically demonstrated benefits to end users.
Model-based UI optimization refers to the use of combinatorial optimization methods to solve a UI design
0 0 1 8 - 9 1 6 2 / 1 7/ $ 3 3 .0 0 © 2 0 1 7 I E E E
problem formulated as a search problem by using predictive models of
human behavior and experience.
The significant benefit of modelbased UI optimization over previous
approaches is that it can find and propose UI designs that are both visually appealing and verifiably usable.
With model-based UI optimization,
a webpage design can be automatically generated so that users regard
its items and colors as aesthetically
pleasing,3 and a programmer can task
an optimizer to design a full menu system with hundreds of commands—
complete with hierarchy, groups, and
42
shortcuts—simply by listing its command names.4
MODEL-BASED UI
OPTIMIZATION
Like a civil engineer whose objective
is building a bridge with a certain
capacity and tolerance, a designer
can create a UI by defining how easily users should find their targets and
how they should perceive it aesthetically. Technically, this is based on the
formulation of a design task mathematically using predictive models of human behavior as objectives.
Figure 1 shows examples of designs
COMPUTER
http://bit.do/anttipaper
generated with recent model-based
approaches to UI optimization,
including application menu design
and menu system design,4 gestural
input design,5 widget GUI design,6
scatterplot design7, and web layout
design.4 Figure 2 shows an overview
of model-based UI optimization from
a design team’s perspective.
UI DESIGN WITH AN
OPTIMIZER IN THE LOOP
Tools might encourage or enforce
user interfaces that were highly
usable, rather than today’s stance
that tools should be neutral… .8
W W W.CO M P U T E R .O R G /CO M P U T E R
Combinatorial design problems
are exceedingly hard for humans
Multiple user groups
Multiple design objectives
Prohibitively large design spaces
A menu with 50 commands can be organized in 10
“A scientific
psychology should
help us in arranging
the interface so it is
easy, efficient and
error free – even
enjoyable.’”
APPROACH OVERVIEW
Optimizer
Generates
Defines task
Steers
Design tool
Designer
Computer scientist
Defines
Evaluates
Behavioral scientist
Predictive
model
Defines
GURE 2. Model-based UI optimization from a designer’s perspective. Predictive models
mathematical or simulations) are used in an optimizer to evaluate generated designs
ainst designer-supplied objectives. The designer can steer and redefine the tasks inter-
COMPUTATIONAL DESIGN
1. DEFINE DECISION PROBLEM
2. DEFINE OBJECTIVES
DESIGN TASK
A COMBINATORIAL DEFINITION
0
1
x1
B x2 C
B C
Find x = B . C 2 X which minimizes f (x)
@ .. A
xn
Multi-objective task
f (x) = !1 f1 (x) + · · · + !q fq (x)
EXAMPLE: LAYOUTS
A B C D E F G H I
J K L M N O P Q R
S T U V W X Y Z
A B C D E F G H I J
A K L M N O P Q R
S T U V W X Y
BACKGROUND: THE LETTER ASSIGNMENT P
Given n letters and n keyslots, the task of the le
OBJECTIVE
FUNCTION
ment
problem is to minimize
the average cost ck`
letter ` after k weighted by the bigram probability
XX
min
p
·
c
k`
k`
A B C D E F G H I
J
K L M N O P Q R
S T U V W X Y Z
Open#
Save##
Save#as...#
Close#
...#
About#
k
`
It is an instance of the Koopmans-Beckman Prob
A B C D E F G H I J
Bigrams
Fitts’
law this prob
which
is
NP-hard.
As
we
will
see
later,
A K L M N O P Q R
S T U V W X Y
modeled with a quadratic function on binary varia
it belongs to a broad class of problems called qu
signment problem (QAP) where the goal is to m
total pair-wise interaction cost [22, 24].
n!##
n!#
“Minimize expected time used for typing in English language”
Except for a few papers using heuristic cost fun
ASSIGNMENT PROBLEM
•
Assign n elements to m slots on a rectangle-sized
surface. Often n = m
Elements
…
Slots
…
DEFINITION OF QAP
Given P facilities and L locations, a weight function w,
and a distance function d, find the bijection
f : P ! L
that minimizes •
X
a,b2P
wa,b df (a),f (b)
w weight
d distance
Intuition: We minimize average distance among
facilities. We try to “pack” together the facilities that
are commonly used with each other
FORMULATION WITH BINARY
DECISION VARIABLES
•
Identify a permulation matrix X of dimensions n x n such that
n
n
X
X
f frequency
min
dih fjk xij xhk
i,j=1 h,k=1
•
d distance
Observations:
•
Large problems (n > 30) may be impossible
•
Generalization of TSP and therefore NP-complete
PRESENTLY ADDRESSED DESIGN PROBLEMS
•
Keyboard layouts
•
Key-target resizing
•
Web pages
•
Menus
•
Mobile apps
•
Dialogues
•
Forms
•
Gestural input
•
•
Layouts
Information
architectures
decreases
more
repetitions.
As experienced
Thedistribution
relationship
between
time
and the
gaze
distribution
Serial
search
refers
to with
thetotal
top-to-bottom
inspection
of items.
users
about
thevisited.
general
location
of users
the target,
Itemshave
aftermemory
the target
are not
With
practice,
not
forEXAMPLE
each
component
is:
MODEL: MENU SELECTION
they
at thebut
neighborhood
of time
the target
and,
evenonlycan
skipglance
more items
also spend less
on each
visited
i=l
X
item. For
this 1
reason,
the number
gaze points
eachsearch,
item
tually,
directly
move the
eyes to of
it [11].
As in on
serial
Tmenu
(✓) =
Gs (i,but
✓)users
+ Gdalso
(i, ✓)
G
✓)of rep(2)
in aassume
isthat
notexperienced
constant
decreases
as
a+function
p (i, items
we
skip
more
and
R
and
d0 the
from of
the
cursor
to
the
when
usingfor
etitions.
The
gaze
points
for gaze
i when
looking
i=0
spend
lessdistance
timenumber
reading
them
than
novices.
t is:
the target
trail strategy.
We assume that d0 = 0.
where
Gs ,model
Gd , G
the numbers
of gaze
points
(seWe thus
distribution
of gaze with
a normal
distrip estimate
⇢this
aspointing)
⇥+exp(
bcaptured
⇥
p
)
+
c
if
i

t
rial,bution
directed,
using
an
eye-tracker
s
i
s
Tand
athe
btarget
⇥ log
(1
+
↵
⇥
t)
(8) i in
centered
on
item
t.
Probability
for
item
p =
2
Gs (i, ✓) =
(3)
with
sampling
rate
ofits
R.probability
In our studies,
R otherwise
= 50Hz. t, 2 (i)
0 by
thea menu
is given
density
function
We can thus estimate Gp . However, to account for the factd
modulated
by
the
power
law
of
learning:
that users can see 2-5 items in the neighborhood of the gaze
Serial search
point, G
wed (i,
revisit
the equation
8.
We
add
a
normal
distribution
2
✓)
=
(i)
⇥
(a
⇥
exp(
b
⇥
p
)
+
c
)
(4)
d
d
i
d
t,
Serial(i)
search
refers
to
d the top-to-bottom inspection of items.
that covers the dispersion of gaze in the items around
t,1.1
Items
after the
target
are
not
visited.
With
practice,
users
not
thewhere
target:
d is the standard deviation reflecting the spread of the
only
skip area.
moreaitems
but
also
spend
less
time
on
each
visited
search
,
b
and
b
are
the
empirical
parameters
of the
d d
d
Gp (i,
✓) reason,
= ap + the
bp ⇥number
log(1 +of
↵ gaze
⇥ t) ⇥
(i) each
(9)item
t,1.1on
item.
For
this
points
learning model. Note that the sum in Eq. 2 ranges from the
in
a
menu
is
not
constant
but
decreases
as
a
function
of
repfirst
item
to
the
end
of
the
menu
(i
2
[1,
l]),
but
not
longer,
[Bailly
et
al.
CHI’14]
where t,1.1 is a normal distribution centered around the tar-
SOME OBJECTIVE FUNCTIONS…
Target selection
Fitts 1954
Internal clock
Lee & Oulasvirta 2016
Saliency
Clutter perception
Itti & Koch 1998
Rosenholtz 2007
Unit perception
Gestalt grouping
Shipley-Kellman 1992 Wertheimer 1938
Reading
Kieras-Hornof 2014 Salvucci 2001
Working memory Multitasking
J.R. Anderson 1998 Wickens 2002
Color harmony
Cohen-Or 2006
Grid quality
Balinsky 2006
Visual search
Strategy adaptation
Chen et al. 2015; Guiard 2015
TOOLBOX FOR DESIGN
PROBLEM-SOLVING WITH OPTIMIZATION
Design tasks
Objective
1.Best possible design?
Des
ig
M. Gilli
2.How far away is the present
design from the best?
n va
riab
le A
Optimization heuristics
Variable B
6
3.Best changes to update a
UI?
4.Best compromise for
multiple user groups
INTEGRATION WITH MACHINE LEARNING
ON-GOING RESEARCH
•
•
•
•
How to steer generated designs to resemble other
designs in this domain? Design mining with statistical
models of designs [CHI’13]
What makes a design good in interaction? Learning
predictive models (e.g., regression models) from data using
Symbolic Programming [CHI’14]
What are my users like? Learning user model parameters
from data using Approximate Bayesian Computation [CHI’17]
How to learn from designs in the wild? Bayesian
optimization and multi-armed bandit approaches to iterative
design improvement in the wild [CHI’16]
CONCLUSION
WHAT CAN I DO TO
COMBINATORIAL UI DESIGN
IS DESIGN FROM FIRST
PRINCIPLES
Models (principles of interaction) allow algorithms to identify
best parameters of a user interface. All inputs are scrutinizable
and empirically verifiable, thus contributing to accumulation of
design knowledge. Integration with data-driven and human-inthe-loop approaches is possible.
Design
Superiority
An
organization
emerges
that can
present
guarantees
for their
designs
Bulk Design
Automated
A design
service
emerges that
use interactive
optimization to
significantly
improve costefficiency in a
popular
category
Computer
Outperforms
Human
SelfDesigning
Interfaces
Computer is
shown to
reliably
outperform
human
designers in
controlled
studies with
end-users
Log data is
used to
parametrize
optimizers that
can optimally
adapt the
interface
without human
intervention.
LEARN MORE!
Internationally
•
CHI’17 course (Denver, May)
•
Summer School (Zurich, June)
•
Book to appear with Oxford Uni Press (August)
Helsinki
•
Self-optimizing Web Services project (email me)
•
HIIT Machine Learning seminar (weekly)
•
Computational UI Design course: students
[email protected] http://bit.do/anttipaper