Spherical Embedding and Classification

18
Spherical Embedding and Classification Richard C. Wilson Edwin R. Hancock Dept. of Computer Science University of York

Transcript of Spherical Embedding and Classification

Spherical Embedding and Classification

Richard C. Wilson

Edwin R. Hancock

Dept. of Computer Science

University of York

Background

Dissimilarities are a common starting point in pattern recognition

• Dissimilarity matrix D

• Find the similarity matrix S

If S is PSD then

– S is a kernel matrix K

– Can use a kernel machine

– Or embed points in a Euclidean space

– And use standard vector-based PR

We can identify D with the (Euclidean) distances between points:

2/1

2/12/1

UX

UUXXKTT

jijjiijijijiij dD xxxxxxxxxxxx ,2,,,),(2

)()(2

1

nn

JID

JIS

Background

Commonly, when comparing structural representations, we do some kind

of alignment

The similarity matrix is usually indefinite (negative eigenvalues) and we

do not get a kernel

There is no representation as points in a Euclidean space

Two basic approaches

– Modify the dissimilarities to make them Euclidean

– Use a non-Euclidean space

Ideally, any non-Euclidean space should be metric and feasible to

compute distances in

0 if , 2/1 i

T UXUUK

Riemannian space

We want to embed the objects in space so we can compute statistics of

them

Need a metric distance measure

Space cannot be Euclidean (normal, flat space)

Riemannian space fulfils all these requirements

• Space is curved – distances are not Euclidean, indefinite similarities

• Metric space

• Distances are measured by geodesics (shortest curve joining two

points in the space)

– Can be difficult to compute the geodesics

– Need to use a space where geodesics are easy

• Obvious choice is space of constant curvature everywhere

Spherical space

• The elliptic manifold has constant positive curvature everywhere

– Can visualise as the surface of a hypersphere embedded in Euclidean

space

– Embedding equation:

– Has positive definite metric tensor, so it is Riemannian

– Sectional curvature is K=1/r2

Well-known example: the sphere

22 rxi

i

yxyxT

dvrvdurds

vrvurvurzyxP

rzyx

,

sin

)cos,sincos,sinsin(),,(

222222

2222

Non-Euclidean geometry

Previous work

• Lindman and Caelli (1978)

– Embedding of psychological similarity data in elliptic and

hyperbolic(negatively curved) space

– Optimisation method suitable for small datasets

• Cox and Cox (1991)

– Define the stress of a configuration on the sphere and minimise

– Difficult optimisation problem

– Not practical on large datasets

• Shavitt and Tankel (2008)

– Embedding of internet connectivity into hyperbolic space

– Physics-based simulation of partical dynamics

In this paper we use the exponential map to develop an

efficient solution for large datasets

Geodesics on the sphere

• A geodesic curve on a manifold is the curve of shortest length joining

two points

– The geodesic distance between two points is the length of the geodesic

joining them

• Geodesics are „great circles‟ of the hypersphere

• Distance between points dependent on the angle and curvature

2

1

2

,cos

cos,

rrd

r

rd

ji

ij

ijji

ijij

xx

xx

ijr

ij

i

j

Elliptic space embedding

Problem:

Find points on the surface of a hypersphere such that the

geodesic distances are given by D

Non-linear constrained optimisation problem

Computationally expensive for large datasets

Our strategy is to update each point position separately

2

1

22*2

,

,cos where

||

min

rrd

r

dd

ji

ij

i

ijijr

xx

x

X

Exponential Map

One of the difficulties of the embedding is that it is constrained to a

hypersphere

The exponential map is a tool from differential geometry which allows us

to map between a manifold and the tangent space at a point

• The tangent plane is a Euclidean subspace

• Log part: from the manifold to the tangent plane

• The exp part goes in the opposite direction

• The map is defined relative to a centre M

YX MLog

XY MExp

M X

Y

Log

Exp

TM

Exponential map for sphere

Map points on the sphere onto the tangent plane

– Map has an origin where sphere touched tangent plane (O)

– Distances (to origin) are preserved (OX=OX‟)

Optimise on the tangent plane

X

X’

Exponential map for the sphere

Given a centre m, a point x on the sphere and a point x‟ on the tangent

plane:

The tangent plane is flat, so distances are Euclidean:

When xi is the centre of the map (xi=m) then the distances are exact,

distortion will occur as xi moves away from the centre

• Project current positions onto tangent plane using xi as centre

• Compute gradient of embedding error on the tangent plane

• Update position xi

sphere) (to 'sin

cos

plane) tangent (to )cos(sin

'

xmx

mxx

)()(2

ij

T

ijijd xxxx

Updating procedure

i

k

i

k

i

j

jiijiji

ji

ijij

E

ddE

ddE

)()1(

2*2

,

22*2

4

xx

xx

For large datasets, computation of second derivatives is

expensive, so we use a simple gradient descent

Can however choose an optimal step size as the smallest root

of the cubic:

)( ,with

023

2*2

222232

ji

T

jijijj

j

jj

j j

jj

j

j

Edd

EEEn

xx

Initialisation

• Need a good initialisation

• Method presented in CVPR10

• If λ0=0 the result is exact

• Otherwise, this a good starting point for our optimisation

2/1

0

*

12

)(minarg

cos)(

ZZ

r

T

rr

rrr

ΛUX

Z

XXDZ

Algorithm

Algorithm

Minimise Z(r) to find initial embedding

Iterate:

For each point xi:

Map points onto tangent plane at xi

Optimise on tangent plane

Map points onto manifold

2/1

0

* ,)(minarg ZZr

rr ΛUXZ

)cos(sin

'

mxx

'sin

cos xmx

i

k

i

k

i

j

jiijiji

E

ddE

)()1(

2*24

xx

xx

Classifiers in Elliptic space

In practical applications, we want to do some kind of learning on the data,

for example classification

• NN classifier is straightforward, as we can compute distances

• In principle, we can implement any geometric classifier as we have a

smooth metric space

– But the classifier must respect the geometry of the manifold

• A simple classifier we can use is the nearest mean classifier (NMC)

– Can compute the generalised mean on the manifold for each class

process) (iterative Log1

Exp

),(minarg

mean dGeneralise

)()(

)1(

2

i

i

k

i

i

kk

n

d

xx

xxx

xx

x

Some examples

*Reproduced from “Classification of silhouettes using contour fragments”

Daliri and Torre, CVIU 113(9) 2009

*

Learning: classification

Conclusions

• We can use Riemannian spaces to represent data from

dissimilarity measures when they cannot be represented in

Euclidean space

• Showed efficient method for embedding in elliptic space

which works on large datasets

– Produces embeddings of low distortion

• Can define simple classifiers which respect the manifold

– NN, NMC

• Need to extend to more sophisticated geometric classifiers