Upload
yuki-koyama
View
4.665
Download
0
Embed Size (px)
Citation preview
Sequential Line Search for Efficient Visual Design Optimization by Crowds
Takeo IgarashiIssei SatoYuki Koyama Daisuke Sakamoto
Motivation
Parameter Tweaking Based on Preference
Design Exploration Optimization
x
⇤= argmax
x2XGoodness(x)
Optimization
x
⇤= argmax
x2XGoodness(x)
Human-in-the-Loop Optimization
Crowdsourced Human Computation
011001011101
Alexander J. Quinn and Benjamin B. Bederson. 2011. Human computation: a survey and taxonomy of a growing field. In Proc. CHI '11. 1403-1412.
Related Work on Crowdsourced Human Computation
Contributions
Contributions
Microtask Design
Microtask Design
😁
Technical Challenges
Technical Background: (Standard) Bayesian Optimization
(Standard) Bayesian Optimization実験計画
New Technique: Bayesian Optimization Based on Line Search
Our Method
comparative な答えしか使えない
How to Define Slider Spaces
S
How to Define Slider Spacesx
+= argmax
x2{xi}µ(x)
x
EI= argmax
x2XEI(x)
Web Interface for Crowdsourcing
Applications #1 Photo Color Enhancement (6D)
Evaluation: Crowdsourced Voting
Applications #2 Material Appearance (3D / 7D)
Comparative Evaluation
Comparative Evaluation
Experiment #1: Synthetic Setting
0.00
0.10
0.20
0.30
0.40
0.50
0.60
0.70
0 5 10 15 20
0.000.050.100.150.200.250.300.350.400.45
0 5 10 15 20
Optimizing a 2D function
Optimizing a 6D function
0.000.100.200.300.400.500.600.700.800.901.00
0 5 10 15 20
0.00
0.10
0.20
0.30
0.40
0.50
0.60
0.70
0.80
0 5 10 15 20
Optimizing a 20D function
0.00
0.20
0.40
0.60
0.80
1.00
1.20
1.40
1.60
0 5 10 15 20
0.00
0.20
0.40
0.60
0.80
1.00
1.20
1.40
0 5 10 15 20
Experiment #2: Crowdsourcing Setting
Experiment #2: Crowdsourcing Setting
0
5
10
15
20
25
30
35
0 2 4 6 8 10 12 14 16 18 20 0
5
10
15
20
25
30
35
0 2 4 6 8 10 12 14 16 18 20
Summary
Summary
Concept:
Strategy:
Technique:
Applications:
Limitation & Future Work
Sequential Line Search for Efficient Visual Design Optimization by Crowds
Takeo IgarashiIssei SatoYuki Koyama Daisuke Sakamoto
Optimization with Different Initial Conditions
Task Burden (Completion Time)
0
10
20
30
40
50
60
70
80
90
SSM 2GC 4GC
Task
Com
plet
ion
Tim
e [s
]
Advantages of Involving Many Crowds
Assumptions on Design Domains
Assumptions on Crowds
Difficult Cases
SIGGRAPH
SIGGRAPH
SIGGRAPH