indicates equal contribution Hao He Dina Katabi Hao Wang ICML 2020 Oral Domain Adaptation One to One Source Domain Target Domain and Many to One Single Target Domain ID: 932095
Download Presentation The PPT/PDF document "Continuously Indexed Domain Adaptation" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Continuously Indexed Domain Adaptation
(* indicates equal contribution)
Hao He*
Dina
Katabi
Hao
Wang
*
ICML 2020 Oral
Slide2Domain Adaptation
One to One
Source Domain
Target Domain
and
Many to One
Single Target Domain
Multiple Source Domains
Many to Many
Multiple Source Domains
Multiple Target Domains
Predict
MNIST
SVHN
RGB
Depth
Problem: Current Domain Adaptation is Categorical
Category 1
Category 2
But many real problem involves continuous domains!
Slide4Medical Applications Require Adaptation Across Age Which is
Continuous
Young
Old
Age
Slide5Self-Driving Car Applications Require Adaptation Across Time Which is
Continuous
Morning
Night
Time
Slide6Example of
Continuously Indexed Domain Adaptation Problems
Domain 1
Domain 2
Domain 15
Domain 30
Setup:
30 continuously indexed domains
Ground-truth labels (
red
and
blue
)
Source:
Domain 1 to 6
Target:
Domain 7 to 30
Task:
Source -> Target
Slide7Performance of Categorical DA
Source
Target
Results
ADDA
:
Adversarial Discriminative Domain Adaptation
(Tzeng, Eric, et al., CVPR 2017)
Slide8Performance of Categorical DA
Source
Target 1
Target 2
Target 3
DANN
: Domain-Adversarial Training of Neural Networks
(
Ganin
,
Yaroslav
, et al., JMLR 2016)
Results
How should we perform
continuous domain adaptation?
Slide9Why Categorical Domain Adaptation Fails
Source:
domain 1
Target:
domain 8
Cannot
capture how the data distribution continuously changes along with the domain index
Target:
domain 15
Distant
Close
Slide10Why Categorial Domain Adaptation Fails
Source:
domain 1
Target:
domain 15
Target:
domain 8
Slide11Solution: Leverage the Domain Index
Source:
domain 1
Target:
domain 15
Target:
domain 8
1
8
15
Slide12Categorical Domain Adaptation
Data
Predicted label
Source or target?
Encoder E
Predictor F
Discriminator D
E
F
D
Slide13Continuously Indexed Domain Adaptation (CIDA)
Data
Domain index
Predicted label
Predicted domain index
Encoder E
Predictor F
Discriminator D
E
F
D
For
:
Minimize
For
,
:
Minimize
Challenge for CIDA
D
omain 1
D
omain 3
D
omain 2
Before training,
Slide15Challenge for CIDA
Ideally, after training
CIDA may end up with
CIDA only matches
To Improve CIDA
Ideally, after training
CIDA may end up with
Continuously Indexed Domain Adaptation (CIDA)
Data
Domain index
Predicted label
Predicted domain index
Encoder E
Predictor F
Discriminator D
E
F
D
For
:
Minimize
For
,
:
Minimize
Probabilistic CIDA (PCIDA)
Data
Domain index
Predicted label
Predicted mean
Encoder E
Predictor F
Discriminator D
For
:
Minimize
Predicted variance
E
F
D
For
,
:
Minimize
PCIDA matches both
and
Theoretical Analysis
Theorem 1
(Informal). CIDA converges,
if and only if
, the expectation of domain index
is identical for any embedding
.
Theorem 2
(Informal). PCIDA converges, if and only if, the expectation and the variance of domain index,
and
are identical for any
.
Theorem 3
(Informal). The global optimum of the two-player game between
and
matches the global optimum of the three-play game between
,
, and
.
Single-Step
Categorical Domain AdaptationADDA: Adversarial Discriminative Domain AdaptationDANN: Domain-Adversarial Training of Neural Networks… (more results in the paper)
Baselines
Multi-Step Categorical Domain Adaptation
CUA : Continuous Unsupervised Adaptation
Domain 1
Domain 2
Domain 3
Domain 4
Slide21Experiments -
Circle
Domain 1
Domain 2
Domain 15
Domain 30
30 continuously indexed domains
Ground-truth labels (
red
and
blue
)
Slide22Experiments -
Circle
The
first 6 domains
as source domains
with the rest as target domains
Ground-truth labels (
red
and blue)
Slide23Experiments -
Circle
DANN
CIDA
CUA
ADDA
Ground truth
Slide24Results (Zoomed in)
Result of applying CIDA
Source
T
arget Part 1
T
arget Part 2
Result of applying CUA
CUA: Local adaptation. Very rugged boundary.
CIDA: Global adaptation. Very
smooth boundary.
Slide25Experiments -
Sine
Similar slides to Circle
Domain 1
Domain 2
Domain 6
Domain 12
12 continuously indexed domains
Ground-truth labels (
red
and
blue
)
Slide26Experiments -
Sine
Similar slides to Circle
Ground-truth labels (
red
and
blue)
The
first 5 domains as source domains
with the rest as target domains
12 continuously indexed domains
Slide27Experiments -
Sine
DANN
CIDA
CUA
ADDA
Ground truth
Slide28Experiments -
Sine
CUA
Domain 5 and 6
Domain 6 and 7
Domain 7 and 8
Domain 8 and 9
Ground truth
CUA may fail in an intermediate domain,
causing failure in all following domains.
Slide29Experiments -
Sine
CIDA
Ground truth
Slide30Real-World Medical Scenario
Sleep Study at Home
Predictor
1. Awake
2. REM
3. Light Sleep
4. Deep Sleep
Input Time-Series Signals:
2. Breathing Belt
1. Nasal Cannula
Slide31Evaluation on Real-World Datasets
SHHS
: Sleep Heart Health StudyMESA
: Multi-Ethnic Study of AtherosclerosisSOF
: Study of Osteoporotic Fractures
Three Sleep Study Datasets
Slide32Continuous Adaptation Setting: Extrapolation
Young
Old
Age
Here ‘age’ is a
domain index
.
Source domains
Slide33Continuous Adaptation Setting: Interpolation
Young
Old
Age
Here ‘age’ is a
domain index
.
Source domains
Source domains
Slide34Results on Extra/Inter-
polation Adaptation
Categorical DA may hurt performance.
Source-only
Categorical DA
Ours
CIDA/PCIDA improve performance in both settings.
CIDA/PCIDA has larger performance gain in extrapolation since it is harder.
Slide35Results on Cross Dataset Adaptation
In the
hardest
transferring tasks (
SOF -> SHHS
/MESA),
PCIDA is a clear winner.
Slide36Multi-Dimensional Domain Index
Domain Index 1: Age
Domain Index 2:
Physical Wellness
Domain Index 3:
Emotional Wellness
Slide37Multi-Dimensional Domain Index
Multi-dimensional indices further improve performance
Multi-dimensional domain index:
Age
Physical wellness
Emotional wellness
Fatigue level
…
Slide38Summary
The first general DA method for continuously indexed domain adaptation.
Theoretical guarantees that CIDA aligns continuously indexed domains at equilibrium.
Two advanced versions, probabilistic CIDA (PCIDA) and
multi-dimensional CIDA.State-of-the-art performance on both synthetic and real-world medical datasets.
Code
will be released soon at https://github.com/hehaodele/CIDA.