tmp881D TMP
tmp881D TMP
Contents
Introduction
Neuro- biological Foundations
1.1
1.2
Retinotopy
1.3
1.4
Results
2.2
.)2
t1
j)(
Stationary points
10
10
t2
13
15
16
Convolution
3.1
18
Discrete 2D-convolution
3.4
4.1 Linearization
4.2 Stability
of stationary points
18
19
20
2L
24
24
25
Numerical simulation
26
5.1
26
5.2 A combination
27
5.3
28
5.4
30
5.5
Sobel-fil1,er
35
35
5.7
36
Conclusion
39
Chapter 1
Neuro-biological Foundations
This Chapter introduces briefly the anatomy of the visual system and the problem
of self-organization in question.
1.1
The level of the receptors lies on the opposite side of the retina, opposite to where
the light comes in (rrinside'r). Between the receptors and the rroutside", we have the
level of the horizontal cells, the bipolar cells, the amacrines and the ganglion cells. The level ofthe receptors consists of2 types ofphoto receptors: 120*106
Stbchen and 6 x 106 Zapfen. The Stbchen are mainly important during times
of low light intensity, whereas the Zapfen are responsible -for color daylight vision.
In the receptor cell there is a molecule called 11-cis-retinal. After absorption of a
photon, the 11-cis-retinal transforms into an isomer, called all-trans-retinal. After
a geometrical change in the relative positions of the molecules, the early receptor
potential is created.
This signal is being transduced to the mentioned cells and finally to the ganglion
cells. The visual information is being transferrpd to the brain via the nervus opticus, which is built out of the axons of the ganglion cells. The nervi optici unite to
the chiasma opticum, where the fibers stemming from the nasal part of the retina
cross to the opposite side. The fibers from the temporal part of the retina stay on
the same side and connect to the first central processing centers of the vision track:
the corpus geniculatum laterale (CGL) and the tectum opticum. Only
about 101,100 of the fibers end at the tectum, the others at the CGL.
The CGL axons t'run" as a fiber called radiatio optica to the primary visual
cortex (area l7). From there, connections exist to the secundary (area 18) and
ternary(area 19) visual cortex; see [ff] .
1.2
Retinotopy
ha^s been
described
in [2, 31. It
Fig.1
Retins
--+++-E
a
U
Modification of weigts. Elevated weights whicl lie closely together iq tle conaection-matrix
cooperxivity (*)- Due to competitior this growth is compensated by a decrease in the
iocrease by
(-).
uses
Adiacent, rieiglrts arnl-,lify tirernst-.lves through cot,r1-relaLiLrtr (*), ancl ilre being inhibi1,ed through comlret,ition trv the corresp,r.,nding rou,s and colun-rns (-) .
1.3
tion
A main feature is the formulation of cooperativity and comlretitivity in mathematical
terms, the main object of investigation is to explain the dynamical betravior of the
weights and to detect general structures.
By anal),tic reasoning several properties of the dynamical systeur modelling the problern aborre can be detected: continuous dependenc5., stationary points, boundedness
of orbits, attractors, . . . . By simulations some questions ma)' be checked concerning
the long-term-behavior of the system. For ttie theory of dl,n3pi.ol systems, see eg.
[16,17,181.
L.4 Results
In this thesis we cornpre tlrr: cyclic convoh tion, r'l here there are r:ilcles of r;ells. rvhich
is biologically unrealistic, u.itli the biologicalll, more plausible linear cornolution (In
[1] the cyrtlir: convolution ri,as used.). Linear convolution means ihat the boundarlcelis are being treated in tlie follou,ing u,ay : tectuur artd retina a,re being continuecl
b1' artificial cells lvith u,eight zero and convolution is done on this artificial fielcl of
c:ells.
Since for linear convolution some anai-vtic steps (unknown stationaly points, invariance of eigenspaces, .. . ) can not be done, a synergetic analysis seems no longer
possible: Therefore we use numerical methods lbr analyzing the dynautics ancl the
Chapter
2.1
[11.
As mentioned above, we assume ihat both retina and tectum are 1-dimensional
chains, each consisting of trf cells. Hence we consider the real (,n[, -n/)-matrix ttt :
(u1,.) of weights between cells t, r . The form of the equations describing the temporal change u' of w is given as follows:
w'r,,:F1,(w')
t,T
:1,..., N,
(2. 1)
wherea)0and
ft,,(r)::a t
tt-4,C6,,(w)
Cr,,(r),:
k(t
t, ,\
- i,t' * i)rui,i
(2 3)
,i,j:|
k(t*i,r-i)utti
where k(t
* i,r - j)
10
{2.4)
The C-operator has been written as a general convolution: we will distinguish between the linear and cyclic convolution (see Chapter 3). The kernel k is normalized
as follows:
N
1.
(2.5)
This property is assumed throughout the thesis without further mention. For most
of the considerations the kernel k is derived from a standard Gaussian:
k(r, :: erp(-(t;'
Usually we put
+ y'1112"'17
a:0.6.
Since the righthand-side F of the differential equation (2.1) is a system of polynomials in the variables arrr it is differentiable and so it is locally Lipschitz continuous.
Hence there exists a uniquely determined local solution for a given initial value 'u.,0:
ru'(s)
F(Tr,'(s)) ,
-s
i0,
7) , tu(0) :
,n0
,uthet'e T >
(2.6)
A;
see [r91.
hypercube
K7,1
:: l0,l/] x .. . x [0,.n/]
is an invariant subset.
Th,en,
Proof:
As we knou,, rve have a uniquely determined local solution
ur :
10,
T)
R.N'N
Suppose
SuTl
,A/
11
.|,,(st) :
tuy,,(s1)Ct,,{*)
- lf2u6lSu, *
Sull
* t)lNc')
-112(N-1)a<0,
-Lf 2w6(s1X(N
This
sho'ws
reasoning, the solution is deflned for all times and stays in the h.vpercube
2.2
knc,rvn
K1r. I
Here we define the irnportant notion of an attractor and determine the attractor
explicitly for the system in question. Notice that the dynamical systern (2.1) is only
meaningfulforbiologicalreasonsrvhenu,0 RI'',: {A : (ar)lau ) 0 for a\li,,i *:
1,...,1/).
\\re clefine the flow induced by the solution ur starting at tuo by:
O(f . uo)
::
tr,(l),
) 0. tr,o e Rf'N
cVlty )0Vu6
i.ff
e VVt>ty@t(t)e
tJ')
[.I
::
[0,
w'r, :
o,(i
0.
U . As
in the proof of
12
j*r
for all s
I c ]R and ut1,(s) I
IY
*2e.
Then
*'rr(t)
cy(1
-.)
+ (N +
z.)iI
k(t
(r/ + dlN)
xl
a(1
Therefore
u(1)
rvhich shows that 'u(s)
s>*efl'L
IL
TL
Therefore, wb(s)
"
e [/ for all
10
Figure 2.1 is an illustration of Theorem 2.3. Three components of the trajectory are
shorvn starting in u(0) : 61[ e IR5'5 ending in a stationary point (from the numerical
point of view) rvhich is approximately equal to the diagonal matrix with entries 4.7
in the diagonal.
2.3
Stationary points
i.fJ
F(P)
13
o.
: n.
c(r)
see
very easily;
Lemma 2.5 Sup,tose that the property C(il) : fl zs satisfi,ed. Then ln the case
a : 0 we haue only one h,omogeneous real stat'ionary po,int, namely the matrir n
uhich'is a matrir w,ith entrtes 1.
Proof:
After some arithmetics,
wl,
1ve
as
N )w 6" [I
u q C 71 (w)]
(2 7)
Therefore o
(*)
r!'i
ll'l*4;
or
a: *J4,
i*,
)Ylfr;:).'
I
Remark 2.6
The reformulatr,on of th,e rigltt-hand .side r,rt, (2.1) suggest t,wo types of
cand'idates of stat'ionary potnts P : (pni) IRN,N;
.
.
pt5
p4
1 euerywhere (P
n)
a,
Definition 2.7 A
Theorem 2.8
lctose permutatr,on
(2 1).
Proof:
See
[101.
t4
0.
2.4
As rve have seen in Section 2.1 the sign of the value of the righthand side
dif{'erent boundaries of the hypercube K1,. is different:
F(...,0,...)r.>0, F'(...,N,...)r<9.
on
(28)
This suggests that F should have a zero P in the hypercube, that is F(P) : 0
due to reasoning as in the mean value theorem. But the mean value theorem is
in a straightfonvard manner applicable in a one dimensional situation only. Fortunately there is a generalization to the higher dimensional case which is knorvn as
the Poincare-\,[iranda Theoreur.; see [20].
{ n and
li,: {r e I" I x1, :0}, If ;: {r : rr :
oTtytos'i,te
1}
faces.
.,
f")
reln: f(t):g
a stat'ionary poi,nt P
in the
for
Prooft
Apply the Theorem in
[20].
Unfortunately, Theorem 2.11 gives no new information when the property C([) : g
is satisfied ,since we know already the stationary point I. But in the linear case,
which is in the focus of this thesis, we have now a stationary point by analytical
reasoning.
15
Definition 2.12
a) An
fo(Xr,..., Xn) :0
where
/, e RlXr,...,,Y,]
,'i:7,. -.,m,
(2 e)
are polynomials.
such that
X, Top
a topolog'ical space.
it a letpologtcal spnce.
Definition 2.15 Let X be a topologi,cal suace.
Th,e Krull rli,ntensi,on dimX of X i,s the supremum of all lengths n of atl cha,ins
* X*t,'i:0,...tzn-r
subsets Xi . Also let dirn0 :: *1
Xo C Xr C
of closed, i,rreduci,ble
"'
C Xn wi,th
Xt.
16
First, let
N:2.
rigthhand-side of our equation. Let V be the corresponding affine varietl,. Let also
A
lvhere
is the ideal
::
in generated b1,the
I ::
assertions:
(a) di:,nV :
1.
impli,es
Conjecture
1:
dimT:1
\,Iost prolrably, it is possible to proof (b) in the Theorern, due to the structure of
the polynomial in cluestion.
dim I,' : 1 means, that the stationary points lbrm an algebraic curve, if cr is considered as a rariable.
If a is a constanl,, then the same conjecture shorvs dim v.- : 0. B), delinition of the
Krrrll dimension, there exist m stationary solutions. nt, ) 7 .
L7
Chapter
Convolution
This chapter gives some defrnitions and results concerning the operation "convolution", mainly because we did not find the discrete 1D, 2D linear convolution
properly defined in the literature. So it should serve as an overview; see [21]. Here
we use the notation which is comrnon in image processing.
3.1
LetnN,n)1.\\reset
'
a.::
exp(2tri,f rt)
r,r,e
have a*.
r'nkrx,J
4,-
k-0
is called
: 0,.... n -
1,
\ :: (tln)\a;ikz1,, j :0,...,n *
-*
C", rvith
7,
k:0
is called inverse
discrete Fourier
tansform (DFT;l).
as
it
is rvell knorvn,
-: -1
3.2
Definition
C.
3.1-
Define
min{L,n-l)
(o * b);
::
a*bt-r,, I :
0, .., 2n
- 2.
k:nraz(0,1-n*1)
a+b
oJ a,b.
1)-tuple over C.
Notation:
Hadamard product: Let z, y tie complex n-tuples.Then r O y is the cornponent'rvise
producl vecLor.
Padding: Let r C', then (rl0;) e qn+l ir the vector rvith I zeros at the end.
linear convolutiou can be computed by using the discrete Fourier transform:
Then
FT;
lD FT2*(a10,)
D FT2"(b10")).
We set
n-1
(a8b)7,: I
a1,bg--x1 mod
n, l:0,. ..,fr-1,
,t:0
convolution of a,b.
Then
FT,(b)1
C.
Then
(a * bl01)
(a10")
19
(bI0").
For a vector J
::
rn-L
t,,
\J"- '
xn*2
lr.
,,\
''n /
Then
40,..,
3.3
)r-r
: Lhr rf^
k:0
Discrete 2D-convolution
min(i,tr-7')
(f
I r,
'r h)(i,
i) ::
k:mat
(0,i.
al*ernet.
mt(j.n-L)
- n*
1) L:mat (0,i * n* r)
tft't)h(i
f (k,l)h(i -- k,',ij - l),i, j :0,
.. ., 2n - 2 ;
n.*1
(f a
h)(t,,
These are
k) -oa ,,(i
respectively.
erp[- (r2 +
a2) I Qo2
)l
h'
Let
aob
:-
20
o,
h'
(u)
..,n
::
erp(-u2 I Qo2))
be the
(a I h'). (b e ') : [I
A:0
n-l
"o
h'(i.
-,t)],
ll,@*b)h(i'
II
bt
h'U
* t)lj
l:0
- k, j - t))u,i :
(a o b)
KI
On the left side is the dyadic product of 1D corvolutions, the riglrt side is a 2D
convolution.
3.4
Now u,e are prepared to define the C-operator by 2D-convolution which u,e warrt to
use in s5-stern(2.1):
Here k
: (ki)
Using Theorem 3.7 (circulant matrix), we can calculate the eigenmatrices and eigenvalues of the cyclic C-operator:
Lemma 3.7O
We hu'ue:
C(y* . a")
fnr all mjn
Proof
:0,
...,ly'
: \*
)'n(y* o y")
C(g*.y") : (y*oy")&k
(a^ A
k,).
(y" A k,)
r
For readabilitl,, su1
rn: n:7
t : (l
N-1
: L
N-1
*;,,'11L nir')
t:0
rr,, erp(N-r2r
i.(p
+ l))
P"L
and
y' .
yt : (rr ,')r,,
erplt{-L2r i(k + t)l
where a,r is the n-th root of unity. These are exactly the expressions obtained by
synergetics in [1].
lrl s
t'
Proof:
Letubeaneigenvector: C)('u)-^tu,ul0.Letlrr,,l:ma-x{lu;3llz,
lrllr,l :
j:1,...,,^,r}.
lc*(r)l
I k(t*i,r*i)uril
i:L,i:t
k(t
i:\,i:l
* i,r * j)lu*|,
k(t-i.r
i.:l,j:l
-.7)
:1.
I
IRN,N,
N.
N:5
?i:
22
l/:9
.^/
7i
7i
15
* . : IR.N,N
---_+ IR.N,N,
l/:5
.
-i
dr
di
N:g
A/:
15
0.9,175
Conjecture
Let 7N
(rfl,".,7nry,)
Then
JIL('N-dN):o'
rvhich rve deline as follows:
MVi:7,...,N2
lryr_,rl
<..
Ji:6i:1.0
23
Chapter 4
Linearization
Now we want to calculate the jacobian Dtr(P) of ttre righthand side f in (2.1) ',vhose
eigenvalues may be used to decide whether a stationarl, point is I'stable'r.
Notation:
Z :: X e Y is the pointu,ise product of X, Y.
Let F(P) : 0 . We linearize F in the neighborhood of the stationary point P. As
in [17] . lVe use
B:IRN'N_-IRN'N,
Br,(W)
.^/,^/
:: (112N)(Lr7 + utj),t,r' :7,.
i:L
.i:r
Using this mapping the righthand-side in (2.1)
F(P)
aI" + P o C(p)
..
,N, tf W : (.0)
* p e B@n+ pe C(p)) :
0.
LetVIRN,N.Then
F(v + P)
: an+ (v+ p) o (cv + cp) * (v +p) o (ar + B((v + p) e (cv + cp)))
: v o cP + P a cv - (y + P) o B(an+v a cv
+vocP+Pocv+PacP)
V oCP + P
oCV
.1
+ p aCV)
Cancelling terms due to F(P) : 0 and neglecting terms which are not linear
obtain for the jacobian of F in P
in V
u,e
DF(P)(v)
: V aCP + P oCV
: v a cP + P ct cv
4.2
Let
fl,
Definition 4.1
Ve
P is called
stable i,ff
> Old > OVr B6(P)(O(t,r) e B,(P) for att r > 0),
ds
P)
iff P
r . The stat,ionar.y
i,s
be a stationary
point.
lent:
a) P is asymptotically stabl,e.
b)
i,s attract,iue.
Proof:
See
[17].
po s,itiu
e and
neg at,iu
e'ig enu
25
Chapter
Numerical simulation
5.1
All computations
are performed
in N,fATLAB.
Our equation (2.1) together rvith an initial condition is an initial value problern
(r\-P):
(5.1)
u' : F(tu), u(0) : uoi
l/
ur.
Let tu be a global sohil,ion. For ttre numerical approximation of this solution \re
the forrvarcl Euler rnethod with step size , :
,!n*1
u(0)
(5 2)
Thus, the approximation of the solution are given by the sequence of values u0 , u1
Notice that each ruk may be considered as matrix in IRN,N.
use
..
:: Mh.
The linear convolution is reaiized b1, using packages in NIe:rLas. Cyclic convolution
is programmed straight{bnvardll, according to the definition.
We shall use pixels for the entries of a matrix trr: the brighter the pixel, the larger
the entry of ir.
diag stands for the rnain, diagx for the diagonal going from left bottom to right
top.
26
Frorn the above, a rather large number of parameters steers each simulation:
l/
a
h
l,t
i,'.:,; ;""*::il:ffi:
o
?uo
q
I ::
10,11 and
q-0:
cyclic
convolution.
Conjecture
3:
that is, only for g : 0 the system shows the cyclic behavior.
If q is only slightly greater than zero, the system switches abruptly into the linear
mode. That is suggested by the follo'iving series of test-calculations:
h:
q: l}-a,\0-7,e:
10*10,r1(o)
: r * --
0.1
> *M a: diag-(4.75)
q:10"2,tu(o) :1-
-*
>
tuL{
xn
14
246a101214
Figure 5.1: The weight matrix wM as a pixel pattern at the end of an iteration.
left:linear,right:cyclic, a : 0.95
5.3
linear
case:
1b.
0.95.
o.os
o.cs
I .o
For a comparison betu,een the linear and cyclic case) see Figure 5.1.
a
For the linear case and tr(O)
wM for
0.93.
0.93, 0.923,0.92
0.5.
Lin,ear case:
1!Lr
:(
0.7377
i3;?'
13.8602
l-,-
0
2B
248101214
2468101214
re 5.2:
Figure
5.2: linear: The transition of wM for a
r
10
29
12
0.92
1t
0.g28 (right)
Cyclic
case:
n.amt o.13oe o
f o'lsos
\
1.3'8s74
o
:l
,.rn,
"-l:::l
0 ... rs.rrzn)
\0
I
cases
0.3, 0.2,0.1
\\'e see the diagonal matrix, for both cases, the diagonal numbers slightly increasing
up to 14.78 i
wM x diag(14.78,.. .,14.78)
;
2466101214
5.4
We simulate the cyclic dynamics, but now with starting point u(0) :: (Aj|) where
A':0 for all i,, j *t,r and Ali: N. All these initial values are loose perutation
matrices.
Let N
:5,a:0.1
and tir(O;
i,1. Then
,rAr
diag(4.7,
30
..
,4.7)
1.5
1.5
la
Figure
0.1., plotted
al',vays
Let
N:5,o:0.1
Let,A/
:5,a:0.1
Finally, let
and
47000 rvhere
,4r1
100000.
and to(O)
ly':5,o:0.1
:43,3.
31
0.5
1
'1.5
2.5
3
-]
1.5
2.5
3.5
4.5
matrix tuM.
2
2.5
J
4
4.5
5
1_5 2
Figure 5.7:
w(0):
2.5
A1,3,
3.5
4.5
32
u.'M
5.5
0.8
0.6
0,4
0.2
0
2.5
A1'3
0.5
2
?.5
3
3.5
4
4.5
5.5
0
,,
Jd
0.5
1
1.5
2
2.5
J
3.5
4
4.5
5
5.5
0
1.5
Figure 5.10:
2.s
T.u(O)
3.5
4.5
0.5
1.5
2
2.5
3
3.5
1.5
2.5
--
34
3.5
4.5
A',t,shown urrl
?.5
2
1.5
0.5
0
5
w(0):
tr+'t.
/t
is
1\
sn:{o o ol
\-r *2 -t)
ur(0): diag(10-2)o:0.1.
5.13:
The rate of convergence is lorver than rvith the Gauss filter for the same other parameters. uLM is stable,r,vith four complex eigenvalues; notice that the Sobel-filter
contains negative entries.
5.6
trO
One aspect of complex systems is the sensitivity with respect to the initial matrix.
This is being illustrated by the nexL two examples. Set -l:5,a:0.1 .
Linear case,
tuo:f
+10-2
35
0.5
1
1.5
2
?.5
3
4
4.5
5
5.5
AA
1.5
2.5
3.5
.M:
4.5
lr
'tD"'
diag(4.75)
try
of
1[.
tuo:1[+i0-2
,nM
diag.(4.75)
5.7
So
The eigenvalues of the linearization in a stationary point are used to decide whether
the stationarl, point is stable.
36
10
Figure 5.14: First perturbation of I.Note that there is hardly any hyperbolic point.
1.4
1.2
.t
0.8
0.6
0.4
0.2
0
2.5
00
J/
Cyclic case :
Let N
5,0
:1[
DF(wM), utilI
0.20
0.20
0.06
y, 'i:
Linear case
l/ :5,a:
Let
**+**
*+***
*****
*
*
*-2.30
u.
seems
to be stable.
stable, too.
3B
1.
diag-(,X.76)
-1.2{)
*2.37
arer gir,en
in the
Conclusion
complex self-organization properties of the Hussler-von
der Malsburg equations are retained when using linear convolution instead of the
original cyclic convolution - which is biologically more realistic. From many starting
points the dynamics reaches one of the 2 diagonal weight matrices. The difference
lies in the algebraic multiplicities of the linearized eigenvalues. This phenomenon
fits exactly in the theory of circulant matrices, which describe the cyclic convolution.
Moreover the dynamics is complex in the sense that tiny changes in the initial
values can produce very different final statioaary states.
Another result is that the biological hypercube K1,, is an attractor for the system.
Also we proved the existence of a stationary solution for the linear convolution,
through the Poincare-Miranda theorem.
We gave also a conjecture about the eigenvalues of the C-operators as l[ goes to
oo. An interesting future investigation wold be the introduction of a model which
describes the system when N goes to oo.
39
?
Bibliography
[1] C. von der Malsburg A. F. Hussler. Development of retinotopic projections:
an analytic treatment. J. Theoret. Neurob'ia., 2:47-73, 1983.
[2] D.J. Willshaw C. von der Malsburg. How to label nerve cells so that they can
interconnect in an ordered fashion. Proc. Natl. Acad. Sci,. USA,74:5L76*5L78,
1977.
[3] D.J. Willshaw C. von der Malsburg. Differential equations for the development
of topological nerve fibre projections. SIAM-AMS Proc.,13:39-47, 1981.
t4l C. von der Nlalsburg. The correlation theory of brain function. MPI Jor Biophysi,cal Chemi,stry, Internal Report
1-2:
1-80, 1981.
[5] C.von der Malsburg J.Zht. Maplets for correspondence-based object recognition. N eural N etw orks, 1 7(B-9) 131 1*1326, 200 4.
:
69
:1305-1308, 2006.
1978.
[Bl H. Haken. Generalized ginsburg-landau equations for phase transition-like phenomena in lasers,non-linear optics, hydrodynamics and chemical reactions. Z.
Physi,k B, 21:105*174, 7975.
L BI Verlag, Mannheim,
1984.
[lS] G.Tnervs.
1982.
[fO1 iV. de 1\Ielo. The Geometry of Dynam,ical Systezns. Springer, Berlin,
4t