• Aucun résultat trouvé

Adaptive architectures for Hebbian network models

N/A
N/A
Protected

Academic year: 2021

Partager "Adaptive architectures for Hebbian network models"

Copied!
11
0
0

Texte intégral

(1)

HAL Id: jpa-00246522

https://hal.archives-ouvertes.fr/jpa-00246522

Submitted on 1 Jan 1992

HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.

L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.

Adaptive architectures for Hebbian network models

Karl Kürten

To cite this version:

Karl Kürten. Adaptive architectures for Hebbian network models. Journal de Physique I, EDP

Sciences, 1992, 2 (5), pp.615-624. �10.1051/jp1:1992105�. �jpa-00246522�

(2)

J. Phys. I France 2

(1992)

615-624 MAY1992, PAGE 615

Classification Physics Abstracts

87.30 75.10H 64.60

Adaptive architectures for Hebbian network models

Karl E. Kfirten

Institut ffir Neuroinformatik, Ruhr-Universitit Bochum, D-4630 Bochum, Germany Institut fur Theoretische Physik, Johannes-Kepler-Universitit Linz, A-4040 Linz, Austria

(Received

13 December 1991, accepted in final form 24

January1992)

Abstract. We present a novel procedure which allows a neural network to evolve to a

quasi-optimal connectivity such that information to be memorized is engraved as efficiently

as possible. The emerging connectivity structure of the resulting partially connected network

depends strongly on the information the network is asked to memorize. A stability measure of the quality of storage as a function of the degree of the connectivity is shown to attain a

maximum value, which lies substantially above the stability of the traditional fully connected system.

I Introduction.

There is no doubt that artificial neural networks which are

supposed

to

perform intelligent

tasks have to be

designed

with great care and considerable

ingenuity. Moreover, tayloring

network architectures

according

to

context-dependent

tasks is a

problem

of fundamental

importance.

Nevertheless,

most of the

currently popular

models are based on fixed network architectures

or even full

connectivity

and lack basic

adaption properties

present in

living

networks. Hence, these models cannot

adapt specific

structures matched to the task the network is asked to

perform.

In contrast, variable

connectivity

for each individual neuron

during

the

learning

session allows the network to

undergo

an

optimization

process, where

optimal

and even minimal

connectivity

structures

might

emerge. Moreover, flexible network

topologies

allow network

designs

or

learning strategies aiming

at

economizing

on the number of connections

highly

matched to the information the network is asked to process.

2. The

partially

connected network model.

Our model system consists of N

binary

formal neurons which assume either the value a; = -I when neuron I is "silent" or a;

= +I when it "fires". In contrast to

Hopfield's

model

ii]

our

system is not

fully

connected and each neuron I receives

input

from

only

It; other units of

(3)

the

network,

which renders the

topology

flexible. The

firing function, h;, representing

the net internal

input

to neuron I at time t is then modelled

by

the

commonly

used linear

superposition

of

weighted input signals

h'(I)

"

LCiJ'J(I)/

II Cl l12

(2.I)

(I) with the

spherical

normalization

II Ci

((2" ~

Cl;

(2.2)

~

The sum

£~;~

is

only

taken over those

K;

neurons that interact with neuron

I,

while self-

coupling

efsects are excluded. The state of the network at time t is then evaluated

according

to the deterministic threshold rule

«;(t

+

I)

=

Sgn(h;(t))

I

= I,

,

N.

(2.3)

Note that the

parallel

or

synchronous updating

in

(2.3)

can be

replaced by

a

sequential

or

asynchronous dynamics

as in the

Hopfield

model

iii.

When a set of p

= aN

arbitrary

memory

configurations presented

as n-Aimensional

firing

patterns

S~,...,Sl'

E

(- I, 1)~

are to be encoded in the

network,

the

oft-diagonal coupling

coefficients can be defined via the Hebbian

prescription

c(

p

=

£ SfS) (2.4)

p=1

usually

called outer

product rule,

or

by

the

corresponding "clipped" coupling

coefficients with

cj

=

sgn(c( (2.5)

Note also that due to the

incomplete connectivity,

the

synaptic couplings

are not

necessarily

correlated for different cells and hence are in

general

no

longer symmetric.

Ideally,

under the

learning

rule

(2A)

and

(2.5)

all the

prescribed

patterns of information should become stable fixed

points

of the

dynamics. However,

it is well known that for

fully

connected networks the choices

(2.4)

and

(2.5)

allow faithful storage of the information

only

up to a critical

storage capacity

ac

= 0,14 and ac

=

0.102, respectively

[2].

By

contrast, we

will see that for

networks,

where each cell chooses its K;

neighbours

in an

optimal

way,

perfect

information

storage

is

possible

far

beyond

these critical

capacities.

3.

Stabflity.

A reliable measure of the

quality

of memorization is

given by

the

magnitude

of the

stability

parameters K;p defined as

K;» =

Sfh;

=

St L c;;SJ/

II C; ll~

(3.i)

u)

In order to store the

prescribed

patterns

S~,...,Sl'

as fixed

points

of the

dynamics (2.3)

it is necessary that all

stability

parameters K;~ be

nonnegative. Furthermore,

their

magnitude

is

required

to be

as

large

as

possible

in order to ensure

reasonably large

basins of attraction for

(4)

N°5 ADAPTIVE ARCHITECTURES FOR HEBBIAN NETWORK MODELS 617

the individual patterns. In order to

gain

more

insight

into the

origin ofstability

we

decompose

K;~, defined for a

fully

connected

network,

into its individual contributions

nj(~

=

Sfc;jS)/

ii c; [[2

I

=

i,.. ,N (3.2)

and define the p x

(N I) stability

matrix

~,

(,(j))j"I,...,N

j#I ~~~~

' ip p=I,...,p

Here,

the matrix element Kj(~

specifies

the

stability

contribution for the I-th component of pattern ~ due to the

synaptic

connection from

j.

Note that

summing

over an

arbitrary

row

j gives merely

£

K)(~ = K;~

,

(3.4)

(I) whereas

summing

over a column ~

produces

f

cj(~

=

~c;jc( /

ii c; [[2

(3.5)

P ~_i P

The sum rule

(3.5) implies

that if and

only

if the actual

couplings

c;j are chosen

according

to the Hebbian

sign

constraint

sgn(c;>)

=

Sgn(c() (3.6)

the arithmetic

stability

mean

4~~ represented by

the left hand side of

equation (3.5)

is strictly

positive,

whereas an anti-Hebbian

sign

restriction

sgn(c;j

=

-sgn(c(

would lead to a

negative

value. This demonstrates

convincingly why

Koehler's and Widmaier's

procedure

[3]

,

based

on

couplings

with the anti-Hebbian

sign restriction,

could not stabilize a

single

pattern of information.

Equation (3.5)

then

implies

that the arithmetic mean of the whole set of the

stability

pa-

rameters Kj(~ taken over all patterns ~ and all

incoming

connections

j

can

always

be written in terms of a normalized sum over the

products

of the actual and the Hebbian

couplings

11 =

L f

c)[~

=

(L c;Jc()/

II C; ll~

(3.7)

u) »=i u)

Hence,

a substantial

portion

of the

couplings

must

obey

the Hebbian

sign

constraint in order to allow for a

positive

mean. Note

however,

that an absolute

stability

measure of the embedded

information is

given by

the value of the minimal ~c;~. In the case of Hebbian

couplings given by equation (2.4), expression (3.2)

reduces to

ii =

)~i

= II

cl

l12

(3.8)

u)

whereas for

"clipped"

Hebbian

couplings specified by equation (2.5), expression (3.7)

takes the form

~ ~

'~'

~~ ~

~~'~

~~

~ ~ ~~ ~~ ~~

(5)

In

judging

the

quality

of memorization one

might

also consider the number of

positive

and

negative

entries in column

j

of the

stability

matrix K; defined in

(3.3), n(

and

n[, respectively.

These numbers

obey

the relations

n(

=

lp

+

c(sgn(c;j

)] and

nj

= §~

c(sgn(c;j

)]

(3.10)

2 2

If the

coupling

c;j is chosen under the Hebbian

sign

constraint

(3.6), n(

assumes its minimum value

)

for

c(

= 0 and reaches its maximum value p for

[c([

= p, where

[c([

takes its maximal

value.

Thus,

the fraction of

positive stability

entries in an

arbitrary

column

j

is

always larger

than and increases

linearly

with

increasing magnitude

of the Hebbian

coupling. Accordingly,

as is

the

case in

biological networks,

it may be inferred that

large

Hebbian efficiencies

play

a

crucial role in

optimized

network models.

4. Dilution of the network

connectivity.

Basically,

there are two

completely difserent, though equivalent approaches

for

reducing

the number of

weights

of a

fully

connected network without

significantly degrading

the

performance

of the network.

Moreover,

even a substantial

improvement

of the

performance

can be achieved

provided

that the

coupling

coefficients have not been chosen

"optimal".

One

might

start from

a

fully

connected network and

selectively

delete those

weights

which disfavour the stabilization of the information to be

engraved.

This

approach

has the serious drawback

that,

before the

dilution process starts, the whole set of

couplings

for a

fully

connected network has to be determined. In

fact,

this

might

be an unfeasible

task,

if nonlocal

learning

rules are taken into account.

By

contrast, we propose to start with a

completely

disconnected network and

selectively

add

only

those connections which are most effective in

increasing

a suitable cost

function,

for

example

the smallest

stability

parameter ~ci~ in

(3.I).

Let us first summarize two

straightforward approaches

in section 4,I and

4.2,

before we propose a

highly

information-

specific

and more efficient

algorithm

in section 4.3.

4.I RANDOM DILUTION. The

simplest strategy

for

reducing

the number of bonds is to

delete bonds at random until the number of synapses for neuron I is reduced to

It;.

The

dynamics

of a

randomly

diluted and hence

asymmetric

version of the

Hopfield

model can be solved

exactly

in the

high-dilution

limit [4], or more

precisely

under the condition It m

logN.

It has been shown that the critical number of unbiased patterns pc the network is able to store is pc

=

~

K.

However,

for an extensive number of

patterns

no

perfect

retrieval is

possible, although~here

exist clouds of attractors near a stored pattern.

Furthermore,

random dilution of a

macroscopic

fraction

synaptic

bonds can also be shown to be

equivalent

to

adding

an

independent

Gaussian noise to

t~e strength

of the

coupling

coefficients.

4.2 DILUTION OF THE SMALLEST BONDS. Just

eliminating

those connections with the

smallest efficacies as

proposed by Morgenstem

[8],

might

be

quite efficient,

but

hardly optimal.

Equation (3.8)

reveals that for fixed

K;

this choice does not

optimize

the minimal K;~ but the arithmetic mean k;. The

price

for such a

straightforward

search may be a

large variance,

whereas a

near-optimal

network structure

requires

a subtle tradeoff between a

larye

average k;

and a small variance. Note that k; reaches its maximum value for a

fully

connected network.

Apparently,

this

quantity

decreases with

increasing degree

of dilution of the weakest

bonds;

therefore, by cutting only those,

the

stability

cannot be

improved

at all.

(6)

N°5 ADAPTIVE ARCHITECTURES FOR HEBBIAN NETWORK MODELS 619

4. 3 SELECTIVE DILUTION.

Conhidering

the

connectivity

parameter K; as

fixed,

the

design

of the network can be

thought

of as a combinatorial

optimization problem,

where each cell I has

(~~)

different ways to choose its

interacting neighbours. Ideally,

one would have to select

j/;

columns of the

stability

matrix defined in

equation (3.8)

such that the minimal

stability

parameter defined in

equation (3.6)

attains its maximum value. Once an

optimal

set of connections has been found for each K;

(I

< K; < N

I),

the

quantity

It; can be used

as a variational parameter in order to determine the

optimal

number of connections which maximize the minimal

stability

parameter K;~. The result is a network with greatest

stability

and minimal resources with respect to the

interconnectivity

structure.

5. Deterministic construction of the network architecture.

We are

mainly

interested in

sparsely

connected

networks,

where the

efficiency

of information

storage

per synapse is

usually

much

higher

than in their

fully

connected counterparts.

Hence,

it is our intention here to build a network connection

by

connection. Since the search process is

independent

for each individual

cell,

we will describe the

algorithm

for a fixed cell I.

In order to attack the combinatorial

optimization problem

we

adapt

Branch-and-Bound

(BILB) algorithms

which have been

widely applied

to the solution of

optimization problems

of

high complexity

[9]. A BILB

algorithm

can be considered as a

general

strategy to search for

optimal

solutions in a

combinatorially large configuration

space, in our case, the space of all

possible

connectivities. This strategy avoids an

explicit

enumeration of all

possible

candidates

by restricting

the search to certain

subspaces

of the whole

connectivity

space. The exclusion of those classes which do not

yield near-optimal

answers from further

competition

is based upon the

computation

of a lower bound of a suitable cost function. To be

precise,

a B&B

algorithm

consists of three components: a

branching scheme,

a

bounding

function and a search strategy.

The

recipe

for

building

our network for a fixed neuron I is a5 follows. We first rearrange N I

possible input

cells

j

in

sequential

order

according

to the number of

positive

entries

n(

defined in

(3.10). (Note

that for Hebbian

couplings

as defined in

(2A)

and

(2.5)

this rearrangement

coincides with

ordering

the cells

according

to the

magnitudes

of their

synaptic weights c;j.)

This ordered set of candidates

(aj~,

aj~,...,aj~-

II

will be called our first

working pool P(I).

Accordingly,

cell

ji

which has the maximal number of

positive

entries

[n(~

[,

provides

the very first connection to cell I. The

correspinding

first lower bound

b(I)

for our BILB-like

algorithm

is then

specified by

the value of the smallest

stability parameter min(K;~)

calculated via

(3,I),

P

where neuron I receives

only

one

single input

from neuron ii such that

b(i)

=

njn(K;»)

=

r§in(SfS[) (5.i)

Notice that the first lower bound

b(I)

will

usually

take the value I, unless there are maximal correlations between the I-th and the

ji-th pixels

among the patterns such that [c;j~ = p and

b(I)

takes the value I. The second neuron which will be connected to neuron I is chosen as follows: Neuron

j2

of our current

working pool P(I),

now

consisting

of

(N 2) units,

serves

as a trial candidate.

According

to

(3.I)

we calculate the trial

stability

measured

by

K~R~j = rrd~

~l(C;J, S(

+

c;j~S()

~

~ (5.2)

corresponding

to two

incoming

connections from neuron

ji

and

j2.

Candidate

j2

is

accepted only

if

K~RaJ >

b(1) (5.3)

(7)

such that the trial

stability

measure

equals

or exceeds the value of the lower bound

b(I).

In

case neuron

j2

is

eligible b(I)

is

replaced by

the

corresponding

value of K,r;ai, otherwise

b(I)

remains

unchanged.

The search strategy continues with a sweep

through

the

remaining

N 3 connections of the

working

pool

P(I

in order to increase further the

stability

of the information the network is to memorize.

Hence,

we accept

only

those neurons, whose trial stabilities

equal

or exceed the current lower bound

b(I).

In other

words, only

neurons which

provide

for an increase of the minimal

stability

will be

eligible.

Let us now assume that the first sweep

provides

neuron I with ni connections from neurons

(ki, k2, ...,kn~)

such that the new

working pool P(2)

consists of N I ni candidates. In

analogy

to the first step we first accept the

neuron with the maximal number of

positive entriesj

regardless

of the

stability

test

(5.3).

The second lower bound

b(2)

is then

specified

with the aid of

(3.I),

where neuron I

now receives ni + I

inputs

from neuron

ki, k2,..

,

kn~+i

km+i

T

b(2)

=

ruin(K;p)

=

min[S,f £ c;iSf]/ £ c[ (5.4)

" "

I=ki I=ki

The search strategy then continues with a second sweep

through

the

working pool P(2)

which

now consists of N 2 ni

remaining

candidates.

In a I-th step, we consider

only

candidates of the

working pool P(I).

In

analogy

to the former steps we first accept the first cell of the ordered set

P(I)

and determine the new lower bound

b(I).

This

quantity

serves as acceptance criterion for

adding

further connections

during

the I-th sweep

through

the

remaining

candidates of

P(I). Again, only

those connections are

accepted

which do not decrease the current lower bond

b(I)

but increase the minimal

stability

measure.

Thus,

after the I-th sweep, we can eliminate another ni connections so that the new

working pool P(I

+

I)

consists of N

£(_~

n; candidates.

This

fully

deterministic BILB-like

procedure

is

repeated

until the

working pool

is empty. It ends after at most

(N I) (weeps,

while the

learning

time is at most of the order of

tJ(N~).

Eventually,

we arrive at a

complete stability

spectrum a5 a function of the

connectivity

and thus can determine those

K;

connections which maximize the smallest

stability

parameter K;p.

However,

it should be

stressed,

that our

algorithm

should be considered a5 a

working

tool for

finding

a

"good"

solution in

polynomial

time rather than a

prescription

for the

"optimal"

solution,

which

might require

a number of steps

exponentially increasing

with the number of

neurons.

6

Computer

simulations.

6.I PROCESSING RANDOM PATTERNS. In this section we demonstrate how the

perfor-

mance of the network varies with

varying connectivity

for three different

strategies

of dilution:

random

dilution,

dilution of the weakest bonds and selective dilution of the

coupling

coeffi- cients as described in the

previous

section. The simulations have been

performed

for N =

100,

N = 400 and N = 800 cells.

Though

we observe some effects of size

dependence

at both ex- tremes of the

degree

of

dilution~

the

qualitative findings

of our

study

are not affected.

Figure

I shows the minimal

stability

parameter Km;n = min K;p,

averaged

over some 50

specimen

nets P

a5 a function of the

degree

of the

connectivity.

The networks consht of N = 100

cells,

while the

capacity

parameter a has been chosen as a

= 0,14. Before the retrieval

phase

the patterns

are

degraded by

random noise and

presented

a5 initial conditions for the network to recall. As to be

expected~

the simulations for random dilution of the

couplings

show a linear decrease

(8)

N°5 ADAPTIVE ARCHITECTURES FOR HEBBIAN NETWORK MODELS 621

2.0

o o

~ l.5 o ° °

~ o ~

2

~

l-O ~

~C

~ ~ ~ ~ * *

~ii 5 ~

n

~

_~ *

~C

~

°

° *

~

fi

n

_5 n

n

~~

o 2 4 6 8 1, o

CONNECTIVITY

Fig. I. Average minimal stability parameter Kmin as a function of the dilution for random dilution

(squares),

dilution of the weakest bonds

(stars)

and selective dilution

(diamonds).

of Kmin, whereas

deleting

bonds with the smallest efficacies

only

leads to a

slight

decrease up to almost 60$l dilution. We observe that the behaviour of Kmin is

quite

reminiscent of recent

exact results for the

capacity

[6].

In the case of selective dilution

following

the scheme of section

3,

we observe that Kmin is a

sharply increasing

function at small

connectivities, reaching

a maximum value

slightly

above 50$l. With further increase of

connectivity,

Kmin shows a slow decrease until it reaches the

limiting

value

corresponding

to the

stability

measure for a

fully

connected

Hopfield

network.

Figure

2

depicts

Kmin for a

= 0A in the case of networks with

"optimal" connectivity.

The

figure

illustrates

clearly that,

over almost the whole

connectivity

range,

coupling

coefficients

"clipped" before

and not

ajier

the

learning

process result in a better

stability

than "

genuine"

Hebbian

coupling

coefficients. Since

clipped couplings

carry less information in

fully

connected

networks,

Kmin takes a lower value for very

high

connectivites.

Preliminary

work shows that a

slightly

modified version of our

neighbour

search

technique

can also be

applied

to the

problem

of ternary

coupling

coefficients c;j E

(0,1, -1)

not restricted

by equation (2.5). Here,

each cell has

exactly (~((~))

different ways to build its

connectivity

structure.

Figure

2 suggests also that for sparse

coniectivity,

networks with

"clipped" couplings perform

as well as the more

general

ternary

choice,

whereas with

increasing connectivity

a substantial part of the nonzero ternary

couplings

do not

obey

the Hebbian

sign

restriction

equation (2.5)

any more.

6. 2 PROCESSING ARTIFICIALLY STRUCTURED INFORMATION. Correlations among differ-

ent patterns turn out to be a serious obstacle for a

satisfactory performance

of the

fully

con- nected

Hopfield

model. This is not

surprising,

since the Hebb rule treats the individual patterns

equally

and considers each pattern as a new

piece

of information.

Thus,

there is a

tendency

to enhance

overlapping

parts of information. On the other

hand, important

details which dis-

tinguish

the patterns are not well

captured by

the network. This failure as well as massive interference erects result in the appearance of numerous

spurious

attractors which deteriorate the network

performance considerably.

(9)

6

~ ~ ,

i

° ° o

~j

2 I n °

[

~

0

~ n u

f~

0 o

u

*

E- ~ n

W -2

_q n *

< -4 a

-6

11

~

-l.0

0 2 .4 .6 .8 1-O

CONNECTIVITY

Fig. 2. Average minimal stability parameter Kmin as a function of the dilution for Hebbiari

(squares),

"clipped"

(stars)

and ternary coupling coefficients

(diamonds).

In real life

applications

stored information is never

completely

random but contains corre- lated structures. For

example,

there are

large

correlations among the 26 letters ofthe

English alphabet

and it is well known that a network model with

couplings

based on the

outer-product

rule

(2.5)

is not able to treat this

problem adequately

in the absence of

high-order

interactions

ill]. However, figure

4 demonstrates that our model is not restricted to capture random infor- mation.

Moreover,

it also shows

good performance

in

processing strongly

structured patterns

interpreted

as different

objects residing

on a uniform

background

as in

figure

3.

iiiiiiiiii lllll'llll ~iiiiiii lll~lll lllllll iiiiii llllll lllll'

lllll~liil iiiiiii::I iiiiiii lJ~~ ff lll~ll

:::::::iii

i~~ ~ ~

8888j°8iI :~~~fl :::::~8fl ~~~fl f%:::ffl

(~~j%ill( (I(li:. ((((1ii8~ H:::::::: ill((((("

fl~::::::: ::::::::~ :::::~flg :::::::::: ff:::@

Fig. 3. Samples the network is to memorize.

Figure

4 shows the recall

performance,

measured

by

the fraction of

recognized bits,

as a

function of the fraction of

correctly presented

bits. It is not

surprising

that the

fully

connected

Hopfield

network

(open squares)

is not able to stabilize this

strongly

correlated

training

set, since the differences in the patterns cannot be well

separated by

the

simple learning

rule and

(10)

N°5 ADAPTIVE ARCHITECTURES FOR HEBBIAN NETWORK MODELS 623

interference effects are too

strong.

Note that

only

six out of the ten

samples

are fixed

points

of

the

dynamics

and hence can be recalled

perfectly. However,

the

"intelligently"

diluted network

(full squares),

where 40$l of the connections have been deleted

according

to our strategy, can

still recall the information almost

completely

even if up to 20$l of the

pixels

have been

damaged.

~

~'~

t

.

n D ° ° ~

~ 9

~

~

~f m

lk- D

~ ~

~ fi

n-

~

~ 7

o

~ .

j~

6

nc

~j

~ ~

5 .6 7 .8 .9 1-o

CORRECT INPUT FRACTION

Fig. 4. Recall performance of a fully connected network

(open squares)

and an optimally diluted network

(full squares).

7. Conclusion and outlook.

We have shown that the

performance

of

selectively

diluted Hebbian network models as mea- sured

by

the minimal

stability

parameter is

superior

to that of the

popular fully connecied

version.

Moreover,

the information is better

stabilized, although

one economizes on the num- ber of

couplings.

In

addition,

our

fully

deterministic

learning strategy

also finds

application

for

binary

and ternary

coupling

coefficients and could also be

applied

to

high-order

networks

[11].

Since the

connectivity

structure of our model system is

highly adapted

to the structure of

the information the network has

learned,

one

might

try to extract

regularities

and correlations from the network

connectivity

a5 well a5 from the

couplings

which have evolved

during

the

learning

session.

Acknowledgements.

The author

acknowledges

support from the

H6chstleistungsrechenzentrum

at KFA Jfilich and from the German Science Foundation under contract number Se

251/32-1.

This work benefitted from

helpful

discussions with J.W.

Clark,

R. Folk~ H. Mfihlenbein and D. Staulfer.

(11)

References

II]

Hopfield J-J-, Proc. Nat. Acad. Sci. 79

(1982)

2554-2558.

[2] van Hemmen J.L., Phys. Rev. A 36

(1987)

1959-1962.

[3] K6hler H-M- and Widmaier D., J. Phys. A 24

(1991)

L495-502.

[4] Kfirten K-E-, J. Phys. France 51

(1990)

1585-1594.

IS] Ktirten K-E-, Parallel Processing in Neural Systems and Computers, R.Eckmiller, G.Hartmann and G. Hauske Eds

(World

Scientific,

1990)

pp.191-194.

[6] Bouten M., Komeda A, and Semeels R., J. Phys. A 23

(1990)

2605-2612.

[7] Derrida B., Gardner E, and Zippelius A., Essrophys. Lett. 4

(1987)

167-173.

[8] Morgenstem I., Lecture Notes in Physics12

(Springer-Verlag,

1986 399-427.

[9] Roucairol C., INRIA report 962,

(1986).

[10] Fontanari J.F, and Theumann W.K., J. Phys. France 51

(1990)

3375-388.

[1ii

Klirten K-E-, Lecture Notes in Physics 368

(Springer-Verlag,

1990 461-466, Statistical Mechanics of Neural Networks, Luis Garrido Ed., XI Sitges Conference on Neural Networks.

Références

Documents relatifs

LOCAL - fully synchronous, messages of unlimited size, each processor acts based on the. information received from up

We first introduce the notion of λ-K¨ onig relational graph model (Definition 6.5), and show that a relational graph model D is extensional and λ-K¨ onig exactly when the

• Models with skill variable with 2 states were better at the very end of tests, but this test stage is is not very important for CAT since the tests usually terminates at early

Countries in the African Region have made more progress over the past 10 years but are still not on track to achieve the health and health-related MDGs despite the

Aware of the grave consequences of substance abuse, the United Nations system, including the World Health Organization, has been deeply involved in many aspects of prevention,

Second, we provide empirical evidence that the choice of an appropriate family of models is often more important—and sometimes much more important, especially when the size of

First introduced by Faddeev and Kashaev [7, 9], the quantum dilogarithm G b (x) and its variants S b (x) and g b (x) play a crucial role in the study of positive representations

To obtain the original data, as a first step it was decided to write software that allows access to the social network through special API and obtain the necessary data from it,