[D2-31] OR

+1 -1 = 1 1 + 2 2 + 3 3

+ 1,2 1 2 + 1,3 1 3 + 2,3 2 3 + 1,2,3 1 2 3

2OR

1 < 1 1 = 0 1,2 = 0 1,2,3 = 0

(), (), (), (), ()

Introduction

Methodology

, ()

Pr in ZnO

{ E1, x1=(0,0,1,0,0,1,0,0,0,1,0,0,0,0,0,0,0,0,1,0) }{ E2, x2=(0,0,0,0,1,0,0,0,0,0,0,0,0,0,0,1,0,0,0,0) }

D2-32

*J. Roh et al., J. Am. Ceram. Soc (2015)

0 : Cu , 1 : Ag

http://www.dokindokin.com/

(D2-33 )

xk+1 = f(xk, uk)

xk

uk

!!

"!

#!!!""

!#!"#

$

% !"#$

$%!&'(!$

=

(D2-33) 1 / 1

D2-34 :

2

CNN adversarialexamples

( ), ( /JSTCREST)

D2-35

clean adversarial

D2-36 () ()

Y G = (Y, E)

q(x) = q(x)/ZZ Bregman-divZ

* Bregman-div p(y |), q(y |)

* Bregman-div

y1By1

y2By2

()

Typeset by FoilTEX 1

le m

a p i

e eee

l bl

0000111000000

m m Undersampling Oversampling / rs

95% n e c

5% n e c

n

l i o Oversampling

e c

e

30min

D2-37

2 () () TOH Kim-Chuan (NUS)

l +

Backtracking (Beck & Teboulle, 09) Modified Backtracking (Scheinberg et al., 14) Restarting (ODonoghue & Candes, 13) unknown Maintaining top-speed (Monteiro et al., 15) unknown

D2-38

O(1/k2)

O(1/k2)

+ O

(log k/k)2

l 2(Takeda, Mitsugi, & Kanamori, 13) SVM

,

AUC

(): , = ,

1 1 1 21 2NEC

A

A B

B

D2-39

!!

!

* NTT

D2'40!

! !! !!

!=!

tt1 t2 t3

!!D2'40!

MDL

2016/11/17 JST CREST Poster Preview

D2-41

(), (), ()

MDL+EM

tt!1t!2t!3t!4t!5

1

3

4

2

D2-42

D2-43SVM

( ), ( )

SVM ( )

m"#$%&'%(

12+

,-+ /,+

[1]

: ,:

10% ,

[1]K.W.ChangandD.Roth.Selectiveblockminimizationforfasterconvergenceoflimitedmemorylarge-scalelinearmodels.In KDD,pages699-707,2011

D2-44 ()

Theoretical Analysis for Parameter Transfer Learning

:

&

D2-45 LPV NEC NEC NEC

LPV

With Luggage State-Space model

Without Luggage State-Space model

LPV

(

D2-46

)(

,(

D2-47Recurrent infomax input-driven recurrent neural network

I(x(t-1); x(t))

x(t)x(t-1)

u(t)

y(t-1) y(t)

u(t-1)

D2-48: Recurrent Neural NetworkNTT ,

RNN (back prop)

l

t(time)

(LSTM, Memory NN) (1+10, 1+5+10, 10)

x

ltj = f

0

@X

i

w

ljix

l1ti +

X

p2P

X

k

(wp)ljk xltpk + b

lj

1

A

Forward prop P = {1, 2}P()

D2-49gamreg:

() ()

IBIS2015

IBIS2016tuning parameterCross-Validation(n=59,p=22,283)

40,000

130R package gamreg

(30)

D2-50: X

Original Image Bicubic Image27.36 dB

SRCNN Image29.55 dB, 500M iterations

(2 days)

9-1-5 3-layer Super Resolution Convolutional Neural Networks (CRCNN)

Computer simulation study- 8000 x 8000 pixel original image- 9000 projections- Binning: 3 in sinogram- FBP reconstruction- No noise added

cf) Dong et. al., ECCV, 2014

D2-51

) ,( ,

(CTMRG)

CTMRG LoopyBeliefPropaga6on

Loopy Belief Propagation CTMRG ( )

NML

LDAD2-52

LDAMDL

MDL: NML

LDA

JST-CREST

2, ,

MDL[Rissanen 1978] NML:

: LDA

, = log , , + log

, ,

+ ()

D2-53 Shapelet

Shapelet

Shapelet

ShapeletShapelet

( )

D2-54: normalized cut ~ NcutSpectral Clustering ~

/ /

(von Luxburg et al. 2008, AoS) Normalized cut (Ncut) Spectral clustering (SC) SC = Q : Ncut (RKHS) Q : Ncut

Spectral Ncut

=

n w k-means

RKHS w k-means

Laplacian matrix

von Luxburg et al. (2008, AoS)

Ncut

Ncut

SC

Dhillon et al. (2007, IEEE PAMI)

Shi and Malik (2000, IEEE PAMI)

n!1!

n!1!6=

**D2-55

min.0 5 10 15 20

opera0

onal

unit

varia

ble

min.0 5 10 15 20

CO2

or

faultm

agnitude

faultm

agnitude

0

10

1

**D2-55

VECRBC

step0 5 10 15 20

step0 5 10 15 20

CO2

opera0

onal

unit

varia

ble

Kriging

Results Al2O310,000

11121 12

D2-56

0

2.0 4.0

dx dz

dy

Introduction

,Varga Istvan

D2-57

Yelp

AUC, , () , /JST(/JST/ThinkCyte), (ThinkCyte)

()

D2-58

MDS MDS()

:98%AUC

AUCCNN

D2-59

, , Preferred Networks

(reparameterization trick)

(x1e-5 x1e-8)

- -Kullback-Leibler

t3 t1

t2

D2-60

x = (x1, x2) y 2 { },

D2-61 AdaGrad

, ( ) ( )

minxRd

{E(a,b)[(ax, b)] + R(x)

}

.

AdaGrad

Lazy Update

D2-62: Shun Watanabe (Tokyo Univ. Agriculture and Technology)A Neyman-Pearson-like Test for Multi-terminal Hypothesis Testing

Xn

Y n

(Xn, Y n) PXY or QXYf1

f2

gzero-rate communication (eg. send marginal types)

Hoeffding-like test [Han-Kobayashi 89]

PXY

P XY

QXY

E(xy(P ))

E(xy(Q))

M(x(P ), y(P ))

Neyman-Pearson-like test

PXY

P XY

QXY

E(xy(P ))

E(xy(Q))

M(x(P ), y(P ))

M(x(Q), y(Q))

QXY

M()

Neyman-Pearson-like test has better error trade-off than Hoeffding-like test for finite block length; Neyman-Pearson-like test has the same (optimal) large deviation performance as Hoeffding-like test; Information geometric observation [Amari-Han 89] plays an important role.

19

c dWD2-63

Wildcard c

u

u

u wildcard ()

NH

NNH

N

NH

S

NS

N

S

O

NO

N

O

wildcard

a

u Wildcard

uwildcard

(!",+1) (!&,1) (!(,+1) (!),+1)

!*+,-.* +1or-1

Active(+1) Inactive(-1)

D2-64 ( )

:

( ), ( ), ( )

1 / 1

D2-65 :

Q : ?

()

raw data k-anonymized data

[A2]

[A2]

0.5 0 0.5 1 1.5 2 2.5 30

0.5

1

1.5

2

2.5

3

Multi

plic

ity

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

2

[A2]

[A2]

0.5 0 0.5 1 1.5 2 2.5 30

0.5

1

1.5

2

2.5

3

Multi

plic

ity

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

2

[A2]

[A2]

0.5 0 0.5 1 1.5 2 2.5 30

0.5

1

1.5

2

2.5

3

Multi

plic

ity0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

2

(), (), ()

105

05

1015

5

0

5

1015

10

5

0

5

10

05

1015

2025

05

1015

20250

5

10

15

20

25

30

05

1015

2025

05

1015

20250

5

10

15

20

25

30

() () ()

D2-66 :