81
IGAIA 4 Bohemia Information Geometry ーー Historical Episodes and Future with Recent Developments Shunichi Amari RIKEN Brain Science Institute

IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

  • Upload
    others

  • View
    4

  • Download
    0

Embed Size (px)

Citation preview

Page 1: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

IGAIA 4    Bohemia

Information Geometryーー Historical Episodes and Future 

with Recent Developments

Shun‐ichi AmariRIKEN Brain Science Institute 

Page 2: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Prehistory ‐‐‐ Riemannian GeometryH. Hotteling 1929

Riemannian metric and Fisher informationlocation‐scale model : constant curvature

P. Ch. Maharanobis 1936    Euclidean distance (multivariate‐Gaussian)

C. R. Rao             1945    Cramer‐Rao Theorem; Riemannian

H. Jeffreys 1946    Bayesian theory and Jeffreys invariant prior

Page 3: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Dual Geometry, InvarianceN. Chentsov 1972 invariance, {g, T}, α‐connection

B. Efron 1975 (A. P. Dawid)   statistical curvature;   higher‐order asymptotics

O. Barndorff‐Nielsen  1976      exponential family; Legendre transform

S. Amari     1982      duality; curvature and statistics (M. Kumon) 

H. Nagaoka and S. Amari   1982     duality, Pythagorean theorem

Page 4: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Amari’s  personal history1958: statistics seminar (master course at U Tokyo)

S. Kullback,  “Information and Statistics”Riemannian metric (suggested by S. Moriguti)

Gausssian : geodesic, constant curvature (Poincare‐half plane)

beautiful structure  essential meaning?mathematical engineering

graph and topology of networks: homologynon‐Riemannian geometry of materials manifold: dislocationsinformation systems, learning and neural networks

2( , )N

Page 5: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Statistical curvature and higher‐order inferenceB. Efron, 1975

Fisher’s idea; exponential connection and mixture connectionA. P. Dawid, 1975  e‐ and m‐connections

S. Amari :     α‐geometry

(Rao, Kano  K?    Fisher’s dream)

Amari and M.Kumon higher‐order power of statistical test

2 1 2 2 22 3

1 1 1( )e m mError G H H Kn n n

Page 6: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Amari paper :  Ann. Statist. 1982: 

Reviewers  (S. Lauritzen and A.P. Dawid)

Chentsov work (handwritten manuscript)

H. Nagaoka and S. Amari  1982 (Technical Report)Ann. Probability Theory      7 reviewersZeitshrift fur Wahrsceinlichkeitstheorie und VerwandteGebietegeometry has nothing to do with statistics

IEEE Trans. Inf. Theory    Shannon Theory, now well‐known

Page 7: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

London Workshop:   1984   (D. Cox)Cox visited Japan in 1983patron of information geometry

Rao, Efron, Dawid, Barndorff‐Nielsen, LauritzenKass, Eguchi many others

Dodson, Critchley, Marriot, Komaki, Zhang, Ay, Pistone, Giblisco, Nielsen, …

Page 8: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Information Geometry ‐‐‐ lucky namingApplications area:statistics, time‐series and systems,machine learning, signal processing, optimizationbrain theory, consciousnessphysics, economics, mathematics 

(Banach manifold, affine differential geometry and beyond) quantum information, Tsallis entropy

Page 9: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

International ConferencesIGAIA series; GSIS series, …Many monographsnew journal    (Jun Zhang);  where to publishmailing list and society

still small community; united and cooperative, blessed by all

Page 10: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

My recent works1. Systems complexity and consciousness (IIT)2. Geometry of score matching (Hyvarinen score)3. Natural gradient descent and topology of deep learning)4. Canonical divergence5. Multi‐terminal statistical inference6. Information geometry and Wasserstein distance

Page 11: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Information Integration and Complexity of Systems

‐‐ Stochastic approach

1x

2x

1y

2y

( , ) ( ) ( | )p p px y x y x

x:  state of the brain      y: next state of the brain

Page 12: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Integrated Information Theory    G. Tononi

Φ

Necessary condition; sufficient?

Page 13: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

full model: Disconnected  model:

measure of interaction  :  N. Ayinformation integration  :  Tononi

Barrett and SethMany other 

1x

2x

1y

2y

1x

2x

1y

2y

{ ( , )}FS p x y

{ ( , )}disS q x y ( | ) ( | )i iq q y x y x

Page 14: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Measure of information integration, or system complexity                

Information Geometry      N. Ay

Page 15: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Definition of : Postulates

1)

2)

3) Disconnected model:  Markov conditions 

min [ ( , ) : ( , )], disqD p x y q x y q S

( , )[ : ] ( , ) log( , )KL

p x yD D p q p x yq x y

Page 16: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Markov Condition

(12) branch deleted: Markov condition:

1 2 2 1 2 2 2( , | ) ( | ) ( | )p x y x p x x p y x1 2 2x x y

: all ( ) deleteddis i jS x y i j

1x

2x

1y

2y

1 2 2

2 1 1

X X YX X Y

Page 17: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Why KL‐divergence? 

1)

2)   

3)

4)

D[p : q] 0, = 0, when and only when p = q

i ii

D[p : q] = d[p(x ), q(x )]

: x

: induces flat structure dually

Page 18: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

min [ ( , ) : ( , )], KL disqgeo D p x y q x y q S

Geometric degree of information integration

Page 19: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Gaussian case

=E[ ]' ' '=E[ ' ' ] ' : diagonal

| ' |log| |

T

T

geo

AA A

y x e eey x e e e

1x

2x

1y

2y

A

Page 20: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Gaussian case

=E[ ]' ' '=E[ ' ' ] ' : diagonal

| ' |log| |

T

T

geo

AA A

y x e eey x e e e

1x

2x

1y

2y

A

Page 21: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Many other definitions of Φ

Full modelDisconnected models

Page 22: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Full model         :  graphical modelFS

12 1 2 12 1 2, exp

higher-order terms

X Y X Y XYi i i i ij i jp x y x x y y x y

x y

, , ,X XYi i ij i jE x E x y

exponential family‐coordinates‐coordinates

x y

There are many disconnected models!!

Page 23: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Split Model            :        Ay,  Barrett & SethHS

12 21 12

,

: min :

ˆ ˆ

0

:

H

S

i i

H KL H KLq S

i iM

XY XY Y

H i i

q q q y x

D p S D p q

q p q p y x

H Y X H

x

Y X

y x

y x p

q

HS

x yx y

1 1 2 2Y X X Y

Page 24: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Mixed Coordinates  :

12 12 21 12 21 12, , , , ; , ,

ˆˆ , ;

X X Y XY XY XY XY Yi i

q p q

x y

12 21 12 ˆ: 0; XY XY YH q qS

x y

Markovian Condition

1 1 2 2

1 2 2 1 1 2

;

Y X X YX X Y Y X X

0, >0Yind Xp p p I x y

problem 0 :p I X Y

Page 25: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Split Model GrS

1 2 2 1 1 2 1 2 2

12 21

1

,

, , , ,

:

ˆ ˆ,

ˆ

0

0

X Y i i

X Y Y Y

i

XY XY

i i i

q q q q y x

q x y x y q x x y q y x y

I X Y

q p q p

q y x p y x

x y x y

x x y y

graphical model

Page 26: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Problem: Gaussian channel( : p A x, y) y x ˆˆ : Gq A S y x

A is not diagonal  12 21 0XY XY

A

:

1 11, exp2 X

p A A

x y x x y x y x

Page 27: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Mismatched Decoding Model MS

Best mismatched decoding from y to x1x

2x

1y

2y

{ ( , ) ( ) ( ) (y | x ) }

* [ : ]

M i i

KL M

S q p p p

D p S

x y x y

ˆy x

Page 28: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

, : , dually flat: , not flat

, ;

H Gr Geo Gr H

M Gr M Geo

Gr H M Geo H

S S S S SS S S S

geo

Page 29: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Transfer Entropy;  Granger causality[ ] min [ : ] ( ) disconnected

[ | ] [ | ]i j KL

j i j

TE x y D p q q i j

H Y X X H Y X

Non‐additive[ ; ] [ ] [ ]i j k m i j k mTE x y x y TE x y TE x y

1x

2x

1y

2y

Page 30: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Hierarchy: transfer entropy

,

,

i i j

i i j

X X X X

Y Y Y Y

cutting branchessplit models

x y

Partition of XX Y

subadditivity

Page 31: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Information Geometry of Hyvärinen Game ScoreFollowing Grinwald, Dawid, Parry, Lauritzen, Hyvärinen

, , :

, , , log

-entropy : ,

-divergence : : :

p x

S p

S S S

L p q E l x a a q x

S x q l x q l x q q x

S H p q E S x q

S D p q H p q H p p

Page 32: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Hyvärinen score

2

2

1, , ,2

1: log log ,2

: :

, , , ,

S p

S x q l x l x

dD p q E p x q xdx

D p cq D p q

x l x l x l x

s

Page 33: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

parametric case { , }:M q A x

min : ,SD p q x

1

, , : estimating function

, 0 : estimating equationN

ii

S

s x x

s x

Information geometry of :SD p q

Page 34: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Asymptotic Analysis of estimator

1 11T TE K VK GN

,

, , T

K E

V E

s x

s x s x

G

Page 35: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

1 1

, ,

, log , ,

T

T

E G AG

A E

c

a x a x

s x x a x

log : efficientq a

Page 36: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Hyvärinen score

2

2

1, , ,2

1: log log ,2

: :

, , , ,

S p

S x q l x l x

dD p q E p x q xdx

D p cq D p q

x l x l x l x

s

Page 37: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Hyvärinen estimator

Fisher efficient q multivariate Gaussian

Page 38: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Discrete case :  graph nodesx

1Nx

f f fN

x

x x x

constxN

x

x

Page 39: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

2

2

,

, 2

:S q

q qS q

q q

q , q ,D E

q , q ,

f h f h

f x h x dx f x h x dx

x

x x

x xx

x x

x xx x

x x x x

, 0f hx

Page 40: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Deep LearningSelf‐Organization  +  Supervised LearningRBM: Restricted Boltzmann MachineAuto‐Encoder,  Recurrent Net

DropoutContrastive divergence

convolution

Page 41: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Mathematical Neurons

i iy w x h w x

x y( )u

u

Page 42: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Multilayer Perceptrons

i iy v w x

, i if v x w x

1 2( , ,..., )nx x x x

1 1( ,..., ; ,..., )m mw w v v

1 w x

yx

Page 43: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Multilayer Perceptron

1 1,

,

, ; ,i i

m m

y f

v

v v

x θ

w x

θ w w

neuromanifold ( )x

space of functions S

Page 44: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Backpropagation --- stochastic gradient learningBackpropagation --- stochastic gradient learning

1 1

2

examples : , , , training set1( , ; ) ,2

log , ;

t ty y

l y x y f

p y

x x

x

x

( , ; )

,

t t tt t

i i

l y x

f v

x w x

Page 45: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

singularities

Page 46: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Geometry of singular model

y v n w x

vv | | 0w

Page 47: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

model: 2 hidden neurons

2

1 1 2 2

2

,

,

12

tu

f w w

y f

u e dt

x J x J x

x

Page 48: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

2

0

1: , ; ,2

: teacher signal :

, ,

t t t

l y y f

y

l y

stochasti

loss fu

backprop : vanilla gradient

c descent learnin

nction

g

 

x x

x

Page 49: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

1 , ,

:

t t t tG y

G l l

Fisher

Natural Gradient Stochastic Descent

Information Matrix

invarint; steepest descent

 

x

Page 50: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Natural gradient is superior

Steepest descent; invariant Yan Ollivier

Fisher‐efficient

Natural gradient is non‐vanishing even in multiple layers

Good at singular regions   (avoid plateaus: Milnor attractor)

Page 51: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Adaptive Natural Gradient

Page 52: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Singular Region in Parameter Space

1 2 1 2

1 2 2

1 2 1

1 1 2 2

, ,

0, ,

, 0,

,

R w w w w

w w w

w w w

f w w

J J J J

J J

J J

x J x J x

Page 53: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Coordinate transformation

1 1 2 2

1 2

1 2

2 1

2 1

1 2

,

,

,

, , ,

w ww w

w w w

w wzw w

w z

J Jv

u J J

v u

Page 54: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Singular Region , 0 1R w z J u

Page 55: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Milnor attractor

Page 56: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Fig. 2: trajectories

Page 57: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Saddle and plateau

Page 58: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Topology of singular R

2 21

2 32

blow-down coordinates , ,

1 ,

1 ,

, 1n

c z u u

c z z u

S

: = e

u

ue eu

Page 59: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Singular Region , 0 1R w z J u

Page 60: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale
Page 61: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Sphere Sn and Projective space Pn

Page 62: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

natural gradient learning near singularity

:

1 :

d Rdt

d O Rdt

 true model

 true model

Milnor attractor

Page 63: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Canonical Divergence in Manifoldof Dual Affine Connections

Nihat Ay and S. Amari

Page 64: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Divergence and metric

3

: 01:2

i jij

D p q

D d g d d O d

: Riemannian metric, positive‐definiteG

Page 65: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Divergence and dual affine connections

:

:

;

ijk ijk

ijk i j k

ijk i j k

i ji j

D

D

Page 66: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Dual geometry

, , ,

, , ,

, , ,

2

X X

ijk ijk ijk

oijk ijk ijk

M g

X Y Z Y Z Y Z

M g T T

T

: Levi‐Civita connectiono

Page 67: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Dual geometry canonical divergence

: dually flat : ,

:

M

D

Bregman divergence

Page 68: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Exponential map :          geodesic t

0

0

1

0 log p

p

q

X q

Xp

q

Page 69: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

p

q

Exponential map divergence

2: :D p q X p q

‐divergence

2: :D p q X p q

‐geodesic

Page 70: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Theorem 1. Exponential map divergenceinduces geometry3

Standard divergence: 2stan 1/3: ,D p q X p q

Theorem 2. exponential mapdivergence recovers the original geometry

13

Page 71: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

1

0 ,

1 2

0 ,

1 2

0 ,

[ : ] ( , ), ( )

|| ( ) ||

( ) || ( ) ||

t q p

q p

q p

D p q X q p t dt

t t dt

w t t dt

ξ

ξ

ξ

Page 72: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Divergence and projection

projection theorem:

ˆ arg min :q S

p D p q

grad :qX c D p q

p

p

S

Page 73: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

IEEE ISIT‐2011 Sankt Petersburg

Data Compression in Multiterminal Statistical Inference

Shun‐ichi AmariRIKEN Brain Science Institute

Page 74: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

A long standing problemT.Berger; Csiszar, Ahlswede, Burnashev, Han, Amari

correlated sources X, Y

data compression and statistical inference1 2 : nX x x x bitsXk

q

1 2 : nY y y y bitsYk

, ; ; p x y q iid

Page 75: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

1binary : , 0,1 ; Prob 1 Prob 12

x y x y

, ; ; p x y q iid Prob x y q

1 2 : nX x x x bitsXkq

1 2 : nY y y y bitsYk

Page 76: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Encoding : data compression

x

nX

c

2nx 2 Xkc

N bits K bits

Page 77: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

One‐bit helper case1, X Yk k n

c

y sgnc a x

1 2 : nX x x x c 1bitq

1 2 : nY y y y n bits

Page 78: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Is single‐bit encoding optimal?

when 1 / 2q

but not for general    .

(         independent), ,x y

q

It is optimal 

Page 79: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Fisher information: 1,X Yk k n

Page 80: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Kingo Kobayashi:  parity encoding

1 2 sx x x

in progress!

Page 81: IGAIA 4 Bohemia Information Geometryigaia.utia.cz/slides/amari.pdf · Prehistory ‐‐‐Riemannian Geometry H. Hotteling 1929 Riemannian metric and Fisher information location‐scale

Information Geometry and Transportation Problem (Wasserstein distance)

entropic relaxation :  min <c, p> ‐ a{‐H(p)}   dual

New Paper

S. Pal and T‐K L. Wong, Exponentially concave function and a new information geometry

Portfolio theory, transportation problem and information geometry(dually projectively flat)