[DL輪読会]Model-Level Dual Learning

DEEP LEARNING JP [DL Papers] Dual Learning Toru Fujino, SCSLAB, UTokyo http://deeplearning.jp/ 1

• http://deeplearning.jp/

n Dual Unsupervised Learning (NIPS 2016) n Dual Supervised Learning (ICML 2017) n Dual Inference for Machine Translation (IJCAI 2017) n Dual Transfer Learning (AAAI 2018) n Model-Level Dual Learning (ICML 2018) n , n n n + +

Duality n n n n n2 (f) (f) (f) (g) (g) (g) !, # n $: & → (, )((|&; -./ ) n 1: ( → &, )(&|(; -/. ) ,

n n n n n etc. n Duality n n n

n Dual Unsupervised Learning (NIPS 2016) n Dual Supervised Learning (ICML 2017) n Dual Inference for Machine for Machine Learning (IJCAI 2017) n Dual Transfer Learning (AAAI 2018) n Model-Level Dual Learning (ICML 2018)

Dual Unsupervised Learning n n 1M ~ n n n

Dual Unsupervised Learning
,

n
!(#|%; '() )

I am a student.

!(%|#; ')( )

,1 1
Unsupervised Learning

n
n
n

n

)

-> dog ->

Dual Unsupervised Learning
n
-(>|?; 9@A )
①
I am a student.
②
-(?|>; 9A@ )

n

!" = $%& (

)

n

!) = log -(I am a student|
n2

! = :!" + 1 − : !)

; 9)

Dual Unsupervised Learning K REINFORCE

Dual Unsupervised Learning n n , -> Dog -> n n n (10%, 100%)

Dual Unsupervised Learning n n Large n Small 100% 10%

Dual Unsupervised Learning n

Dual Unsupervised Learning , Unsupervised Learning n n n n Future Work

n Dual Unsupervised Learning (NIPS 2016) n Dual Supervised Learning (ICML 2017) n Dual Inference for Machine for Machine Learning (IJCAI 2017) n Dual Transfer Learning (AAAI 2018) n Model-Level Dual Learning (ICML 2018)

Dual Supervised Learning n Primal task, Dual task ! " #; %&' !(#|"; %'& ) , n ! " #; %&' = ! # "; %'& = ! #, " for any #, " n n

Dual Supervised Learning
n

!"#$%&'( = log - . + log - 0 .; 23( − log - 0 − log - . 0; 2(3
n

task
7
ℒ 23( = 8
∑[!7 ; .< ; 23( , 0< + >3( !"#$%&'( (.< , 0< ; 23( , 2(3 )]
7
ℒ 2(3 = 8
∑[!7 ; .< ; 23( , 0< + >3( !"#$%&'( (.< , 0< ; 23( , 2(3 )]

5

Dual Supervised Learning nf g

Dual Supervised Learning n n n

Dual Supervised Learning

Dual Supervised Learning

Dual Supervised Learning n

n Dual Unsupervised Learning (NIPS 2016) n Dual Supervised Learning (ICML 2017) n Dual Inference for Machine for Machine Learning (IJCAI 2017) n Dual Transfer Learning (AAAI 2018) n Model-Level Dual Learning (ICML 2018)

Dual Inference for Machine Learning n Primal task, Dual task !: # → % &: % → # '( ), + , ', (), +) n , : + ⋆ = 12&max ' ), + ( 7 )⋆ n! + ) n ) 6 ∈9 : , +) = 12& min ' () , 7 = ∈> ,& +

Dual Inference for Machine Learning
,

n

(Dual Inference)
3 + 1 − . / (1, ! 3 )
! ⋆ = $%& min
./
1,
!
0
7
+
* ∈-

3
3
&⋆ = $%& min
</
1
,
!
+
1
−
<
/
(1
, !)
7
0
+
: ∈;

Dual Inference for Machine Learning Dual Inference n 1. ! K 2. K "#$ %, (' (*) + 1 − " #/ (%, (' * )

Dual Inference for Machine Learning n Dual Inference /

Dual Inference for Machine Learning n (α, β)

Dual Inference for Machine Learning n

Dual Inference for Machine Learning n

n Dual Unsupervised Learning (NIPS 2016) n Dual Supervised Learning (ICML 2017) n Dual Inference for Machine for Machine Learning (IJCAI 2017) n Dual Transfer Learning (AAAI 2018) n Model-Level Dual Learning (ICML 2018)

Dual Transfer Learning n !(#|%; '() ) , n ! # = , ! # %; '() !(%) (∈. n n

Dual Transfer Learning (AAAI 2018) , n # $ = ∑# $ '; )*+ #(') ≈ 0/ ∑#($|' 2 ; )*+ ), n # $ '; )*+ , ' ∼ #(') #('|$) # $ = ∑# $ '; )*+ #(') ≈ 0/ ∑ n #('|$; )+* ) n ' ( # $ ' 2 ; )*+ # ' # ' 2 $ 2 ∼ #(') ) , 2 , ' 2 ∼ #('|$; )+* )

Dual Transfer Learning n n! "

Dual Transfer Learning n

Dual Transfer Learning n

Dual Transfer Learning n λ

Dual Transfer Learning n !(#|%; '() )

Dual Transfer Learning n n n

n Improving Neural Machine Translation Models with Monolingual n n n n Data (ACL 2016) Neural Machine Translation with Reconstruction (AAAI 2017) Iterative Back-Translation for Neural Machine Translation (WNMT 2018) Joint Training for Neural Machine Translation Models with Monolingual Data (AAAI 2018) Unsupervised Machine Translation Using Monolingual Corpora Only (ICLR 2018)

n n n n n n n n Dual Unsupervised Learning Dual Inference Dual Transfer Learning + Dual Supervised Learning Model-Level Dual Learning