simultaneous,deep,transfer,across, domains,and,tasks
TRANSCRIPT
Simultaneous,Deep,Transfer,Across,Domains,and,Tasks,
!Presenta)on!by!Alejandro,Cartas!
Eric!Tzeng,!Judy!Hoffman,!Trevor!Darrell,!Kate!Saenko!
Domain Adaptation: Train on source adapt to target
backpack chair bike
Source Domainlots of labeled data
⇠ PS(X,Y )
DS = {(xi, yi), 8i 2 {1, . . . , N}}
Domain Adaptation: Train on source adapt to target
backpack chair bike
Source Domainlots of labeled data
⇠ PS(X,Y )
DS = {(xi, yi), 8i 2 {1, . . . , N}}
bike??
Target Domainunlabeled or limited labels
⇠ PT (Z,H)
?DT = {(zj , ), 8j 2 {1, . . . ,M}}
Domain Adaptation: Train on source adapt to target
backpack chair bike
Adapt
Source Domainlots of labeled data
⇠ PS(X,Y )
DS = {(xi, yi), 8i 2 {1, . . . , N}}
bike??
Target Domainunlabeled or limited labels
⇠ PT (Z,H)
?DT = {(zj , ), 8j 2 {1, . . . ,M}}
Source Data
backpack chair bike
fc8conv1 conv5source data
fc6 fc7 classif cationloss
i
Adap)ng!across!domains!minimize!discrepancy!!
L(xS , yS , xT , yT , ✓D; ✓repr, ✓C) =LC(xS , yS , xT , yT ; ✓repr, ✓C)
+�Lconf
(xS
, x
T
, ✓
D
; ✓repr
)
+⌫Lsoft
(xT
, y
T
; ✓repr
, ✓
C
)
Eric!Tzeng,!et.!al.!Sim
ultane
ous!D
eep!Transfer!Across!D
omains!and
!Tasks,!2015!
Adap)ng!across!domains!minimize!discrepancy!!
L(xS , yS , xT , yT , ✓D; ✓repr, ✓C) =LC(xS , yS , xT , yT ; ✓repr, ✓C)
+�Lconf
(xS
, x
T
, ✓
D
; ✓repr
)
+⌫Lsoft
(xT
, y
T
; ✓repr
, ✓
C
)
Source Data
backpack chair bike
Target Databackpack
?
fc8conv1 conv5 fc6 fc7
labeled target data
fc8conv1 conv5source data
fc6 fc7
classif cationlosssh
ared
shar
ed
shar
ed
shar
ed i
Eric!Tzeng,!et.!al.!Sim
ultane
ous!D
eep!Transfer!Across!D
omains!and
!Tasks,!2015!
Adap)ng!across!domains!minimize!discrepancy!!LD(xS , xT , ✓repr; ✓D) = �
X
d
[yD = d] log qd Lconf
(x
S
, x
T
, ✓
D
; ✓
repr
) = �X
d
1
D
log q
d
Adapted!from!J.!Hoffman,!Adap)ng!Deep!Networks!Across!Domains,!Modali)es,!and!Tasks,!2015!
q = softmax(✓TDf(x; ✓repr))
Adap)ng!across!domains!minimize!discrepancy!!
✓Cobjectclassifier
LD(xS , xT , ✓repr; ✓D) = �X
d
[yD = d] log qd Lconf
(x
S
, x
T
, ✓
D
; ✓
repr
) = �X
d
1
D
log q
d
q = softmax(✓TDf(x; ✓repr))
Adapted!from!J.!Hoffman,!Adap)ng!Deep!Networks!Across!Domains,!Modali)es,!and!Tasks,!2015!
Adap)ng!across!domains!minimize!discrepancy!!
✓Cobjectclassifier
LD(xS , xT , ✓repr; ✓D) = �X
d
[yD = d] log qd Lconf
(x
S
, x
T
, ✓
D
; ✓
repr
) = �X
d
1
D
log q
d
q = softmax(✓TDf(x; ✓repr))
Adapted!from!J.!Hoffman,!Adap)ng!Deep!Networks!Across!Domains,!Modali)es,!and!Tasks,!2015!
Adap)ng!across!domains!minimize!discrepancy!!
✓Cobjectclassifier
Discrepancy
LD(xS , xT , ✓repr; ✓D) = �X
d
[yD = d] log qd Lconf
(x
S
, x
T
, ✓
D
; ✓
repr
) = �X
d
1
D
log q
d
q = softmax(✓TDf(x; ✓repr))
Adapted!from!J.!Hoffman,!Adap)ng!Deep!Networks!Across!Domains,!Modali)es,!and!Tasks,!2015!
Adap)ng!across!domains!minimize!discrepancy!!
✓Cobjectclassifier domain
classifier
✓D
LD(xS , xT , ✓repr; ✓D) = �X
d
[yD = d] log qd Lconf
(x
S
, x
T
, ✓
D
; ✓
repr
) = �X
d
1
D
log q
d
q = softmax(✓TDf(x; ✓repr))
Adapted!from!J.!Hoffm
an,!Adap)ng!Deep!Netw
orks!Across!Domains,!M
odali)es,!and!Tasks,!2015!
Adap)ng!across!domains!minimize!discrepancy!!LD(xS , xT , ✓repr; ✓D) = �
X
d
[yD = d] log qd Lconf
(x
S
, x
T
, ✓
D
; ✓
repr
) = �X
d
1
D
log q
d
domainclassifier
✓D
q = softmax(✓TDf(x; ✓repr))
Adapted!from!J.!Hoffm
an,!Adap)ng!Deep!Netw
orks!Across!Domains,!M
odali)es,!and!Tasks,!2015!
Adap)ng!across!domains!minimize!discrepancy!!LD(xS , xT , ✓repr; ✓D) = �
X
d
[yD = d] log qd Lconf
(x
S
, x
T
, ✓
D
; ✓
repr
) = �X
d
1
D
log q
d
q = softmax(✓TDf(x; ✓repr))
✓Cobjectclassifier
Adapted!from!J.!Hoffm
an,!Adap)ng!Deep!Netw
orks!Across!Domains,!M
odali)es,!and!Tasks,!2015!
Adap)ng!across!domains!minimize!discrepancy!!
L(xS , yS , xT , yT , ✓D; ✓repr, ✓C) =LC(xS , yS , xT , yT ; ✓repr, ✓C)
+�Lconf
(xS
, x
T
, ✓
D
; ✓repr
)
+⌫Lsoft
(xT
, y
T
; ✓repr
, ✓
C
)
Source Data
backpack chair bike
Target Databackpack
?
fc8conv1 conv5 fc6 fc7 all t
arge
t dat
a
source data
labeled target data
fc8conv1 conv5source data
fcD
fc6 fc7
classif cationloss
domainconfusion
loss
domainclassif er
losssh
ared
shar
ed
shar
ed
shar
ed
shar
ed
i
i
Eric!Tzeng,!et.!al.!Sim
ultane
ous!D
eep!Transfer!Across!D
omains!and
!Tasks,!2015!
Domain Adaptation: Train on source adapt to target
backpack chair bike
Adapt
Source Domainlots of labeled data
⇠ PS(X,Y )
DS = {(xi, yi), 8i 2 {1, . . . , N}}
bike??
Target Domainunlabeled or limited labels
⇠ PT (Z,H)
?DT = {(zj , ), 8j 2 {1, . . . ,M}}
Domain Adaptation: Train on source adapt to target
backpack chair bike
Adapt
Source Domainlots of labeled data
⇠ PS(X,Y )
DS = {(xi, yi), 8i 2 {1, . . . , N}}
bike??
Target Domainunlabeled or limited labels
⇠ PT (Z,H)
?DT = {(zj , ), 8j 2 {1, . . . ,M}}
Source!soKlabels!
SourceCNN
SourceCNN
SourceCNN
BottleMug Chai
rLapt
opKeyb
oard
Bottle Mug Chai
rLapt
opKeyb
oard
Bottle Mug Chai
rLapt
opKeyb
oard
Bottle Mug Chai
rLapt
opKeyb
oard
+
softmaxhightemp
softmaxhightemp
softmaxhightemp
Eric!Tzeng,!et.!al.!Sim
ultane
ous!D
eep!Transfer!Across!D
omains!and
!Tasks,!2015!
l(bottle)
Source!soKlabels!
Bottle Mug ChairLaptop
Keyboard
Bottle Mug ChairLaptop
Keyboard
Adapt CNN
“Bottle”
Source ActivationsPer Class
backprop
Cross Entropy Loss
softmaxhightemp
Eric!Tzeng,!et.!al.!Sim
ultane
ous!D
eep!Transfer!Across!D
omains!and
!Tasks,!2015!
Lsoft
(x
T
, y
T
; ✓
repr
, ✓
C
) = �X
i
l
yTi
log p
i
p = softmax(✓TCf(xT ; ✓repr)/⌧)
Class!correla)on!transfer!loss!
L(xS , yS , xT , yT , ✓D; ✓repr, ✓C) =LC(xS , yS , xT , yT ; ✓repr, ✓C)
+�Lconf
(xS
, x
T
, ✓
D
; ✓repr
)
+⌫Lsoft
(xT
, y
T
; ✓repr
, ✓
C
)
Source Data
backpack chair bike
Target Databackpack
?
fc8conv1 conv5 fc6 fc7
Source softlabels
all t
arge
t dat
a
source data
labeled target data
fc8conv1 conv5source data
softmaxhigh temp
softlabelloss
fcD
fc6 fc7
classif cationloss
domainconfusion
loss
domainclassif er
losssh
ared
shar
ed
shar
ed
shar
ed
shar
ed
i
i
Eric!Tzeng,!et.!al.!Sim
ultane
ous!D
eep!Transfer!Across!D
omains!and
!Tasks,!2015!
Office dataset Experiment Adapting Visual Category Models to New Domains 9
31 categories� �� �
keyboardheadphonesfile cabinet... laptop letter tray ...
amazon dSLR webcam
...
inst
ance
1in
stan
ce 2
...
...
...
inst
ance
5
...
inst
ance
1in
stan
ce 2
...
...
...
inst
ance
5
...
� �� �3 domains
Fig. 4. New dataset for investigating domain shifts in visual category recognition tasks.Images of objects from 31 categories are downloaded from the web as well as capturedby a high definition and a low definition camera.
popular way to acquire data, as it allows for easy access to large amounts ofdata that lends itself to learning category models. These images are of productsshot at medium resolution typically taken in an environment with studio lightingconditions. We collected two datasets: amazon contains 31 categories4 with anaverage of 90 images each. The images capture the large intra-class variation ofthese categories, but typically show the objects only from a canonical viewpoint.amazonINS contains 17 object instances (e.g. can of Taster’s Choice instantco↵ee) with an average of two images each.
Images from a digital SLR camera: The second domain consists of im-ages that are captured with a digital SLR camera in realistic environments withnatural lighting conditions. The images have high resolution (4288x2848) andlow noise. We have recorded two datasets: dslr has images of the 31 object cat-
4 The 31 categories in the database are: backpack, bike, bike helmet, bookcase, bottle,calculator, desk chair, desk lamp, computer, file cabinet, headphones, keyboard, lap-top, letter tray, mobile phone, monitor, mouse, mug, notebook, pen, phone, printer,projector, puncher, ring binder, ruler, scissors, speaker, stapler, tape, and trash can.
• all classes have source labeled examples
• 15 classes have target labeled examples
• evaluate on remaining 16 classes
[saenko`10]
540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593
594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647
ICCV#937
ICCV#937
ICCV 2015 Submission #937. CONFIDENTIAL REVIEW COPY. DO NOT DISTRIBUTE.
A ! W A ! D D ! A D ! W W ! A W ! D Average
DLID [7] 51.9 – – 78.2 – 89.9 –DeCAF6 S+T [9] 80.7 ± 2.3 – – 94.8 ± 1.2 – – –DaNN [13] 53.6 ± 0.2 – – 71.2 ± 0.0 – 83.5 ± 0.0 –Source CNN 56.5 ± 0.3 64.6 ± 0.4 47.6 ± 0.1 92.4 ± 0.3 42.7 ± 0.1 93.6 ± 0.2 66.22Target CNN 80.5 ± 0.5 81.8 ± 1.0 59.9 ± 0.3 80.5 ± 0.5 59.9 ± 0.3 81.8 ± 1.0 74.05Source+Target CNN 82.5 ± 0.9 85.2 ± 1.1 65.8 ± 0.5 93.9 ± 0.5 65.2 ± 0.7 96.3 ± 0.5 81.50
Ours: dom confusion only 82.8 ± 0.9 85.9 ± 1.1 66.2 ± 0.4 95.6 ± 0.4 64.9 ± 0.5 97.5 ± 0.2 82.13Ours: soft labels only 82.7 ± 0.7 84.9 ± 1.2 66.0 ± 0.5 95.9 ± 0.6 65.2 ± 0.6 98.3 ± 0.3 82.17Ours: dom confusion+soft labels 82.7 ± 0.8 86.1 ± 1.2 66.2 ± 0.3 95.7 ± 0.5 65.0 ± 0.5 97.6 ± 0.2 82.22
Table 1. Multi-class accuracy evaluation on the standard supervised adaptation setting with the Office dataset. We evaluate on all 31 categoriesusing the standard experimental protocol from [28]. Here, we compare against three state-of-the-art domain adaptation methods as well as aCNN trained using only source data, only target data, or both source and target data together.
A ! W A ! D D ! A D ! W W ! A W ! D Average
MMDT [18] – 44.6 ± 0.3 – – – 58.3 ± 0.5 –Source CNN 54.2 ± 0.6 63.2 ± 0.4 36.4 ± 0.1 89.3 ± 0.5 34.7 ± 0.1 94.5 ± 0.2 62.0
Ours: dom confusion only 55.2 ± 0.6 63.7 ± 0.9 41.2 ± 0.1 91.3 ± 0.4 41.1 ± 0.0 96.5 ± 0.1 64.8Ours: soft labels only 56.8 ± 0.4 65.2 ± 0.9 41.7 ± 0.3 89.6 ± 0.1 38.8 ± 0.4 96.5 ± 0.2 64.8Ours: dom confusion+soft labels 59.3 ±0.6 68.0±0.5 43.1± 0.2 90.0± 0.2 40.5±0.2 97.5± 0.1 66.4
Table 2. Multi-class accuracy evaluation on the standard semi-supervised adaptation setting with the Office dataset. We evaluate on 16held-out categories for which we have no access to target labeled data. We show results on these unsupervised categories for the source onlymodel, our model trained using only soft labels for the 15 auxiliary categories, and finally using domain confusion together with soft labelson the 15 auxiliary categories.
target domain. We report accuracies on the remaining un-labeled images, following the standard protocol introducedwith the dataset [28]. In addition to a variety of baselines, wereport numbers for both soft label fine-tuning alone as wellas soft labels with domain confusion in Table 1. Because theOffice dataset is imbalanced, we report multi-class accura-cies, which are obtained by computing per-class accuraciesindependently, then averaging over all 31 categories.
We see that fine-tuning with soft labels or domain con-fusion provides a consistent improvement over hard labeltraining in 5 of 6 shifts. Combining soft labels with do-main confusion produces marginally higher performance onaverage. This result follows the intuitive notion that whenenough target labeled examples are present, directly opti-mizing for the joint source and target classification objective(Source+Target CNN) is a strong baseline and so using ei-ther of our new losses adds enough regularization to improveperformance.
Next, we experiment with the semi-supervised adaptationsetting. We consider the case in which training data andlabels are available for some, but not all of the categories inthe target domain. We are interested in seeing whether wecan transfer information learned from the labeled classes tothe unlabeled classes.
To do this, we consider having 10 target labeled exam-ples per category from only 15 of the 31 total categories,
following the standard protocol introduced with the Officedataset [28]. We then evaluate our classification performanceon the remaining 16 categories for which no data was avail-able at training time.
In Table 2 we present multi-class accuracies over the 16held-out categories and compare our method to a previousdomain adaptation method [18] as well as a source-onlytrained CNN. Note that, since the performance here is com-puted over only a subset of the categories in the dataset, thenumbers in this table should not be directly compared to thesupervised setting in Table 1.
We find that all variations of our method (only soft labelloss, only domain confusion, and both together) outperformthe baselines. Contrary to the fully supervised case, here wenote that both domain confusion and soft labels contributesignificantly to the overall performance improvement of ourmethod. This stems from the fact that we are now evaluat-ing on categories which lack labeled target data, and thusthe network can not implicitly enforce domain invariancethrough the classification objective alone. Separately, thefact that we get improvement from the soft label training onrelated tasks indicates that information is being effectivelytransferred between tasks.
In Figure 5, we show examples for theAmazon!Webcam shift where our method correctlyclassifies images from held out object categories and the
6
Office dataset Experiment
Multiclass accuracy over 16 classes which lack target labels
back packbike
bike helmet
bookcasebottle
calculator
desk chair
desk lamp
desktop co
mputer
file ca
binet
headphones
keyboard
laptop computer
letter tray
mobile phonemonitor
mousemug
paper notebookpenphone
printer
projector
punchers
ring binderruler
scissorsspeaker
stapler
tape dispenser
trash can
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.1Ours soft label
back packbike
bike helmet
bookcasebottle
calculator
desk chair
desk lamp
desktop co
mputer
file ca
binet
headphones
keyboard
laptop computer
letter tray
mobile phonemonitor
mousemug
paper notebookpenphone
printer
projector
punchers
ring binderruler
scissorsspeaker
stapler
tape dispenser
trash can
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.1Baseline soft label
ring bindermonitor
Baseline soft activation
Our soft activation
Target test image
back pack bike bike helmet
bookcase bottle calculator
desk chair desk lamp desktop computer
file cabinet headphones keyboard
laptop computer letter tray mobile phone
Source soft labels
Cross-dataset Experiment Setup
Source: ImageNet !Target: Caltech256 !40 categories !Evaluate adaptation performance with 0,1,3,5 target labeled examples per class
[tommasi`14]
ImageNet adapted to Caltech
Number Labeled Target Examples per Category0 1 3 5
Mul
ti-cla
ss A
ccur
acy
72
73
74
75
76
77
78
Source+Target CNNOurs: softlabels onlyOurs: dom confusion+softlabels
[ICCV 2015]
400 120 200Number of labeled target examples
Mul
ticla
ss A
ccur
acy