Int
ern
at
i
onal
Journ
al of Ele
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
8
,
No.
6
,
D
ece
m
ber
201
8
, pp.
4292
~
43
09
IS
S
N:
20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v
8
i
6
.
pp
4292
-
43
09
4292
Journ
al
h
om
e
page
:
http:
//
ia
es
core
.c
om/
journa
ls
/i
ndex.
ph
p/IJECE
Cereb
ell
ar M
odel Cont
ro
ller with
new Mo
del of
Gr
anule
Cell
-
g
olg
i Cell Bu
ild
ing Blocks
and
Two
-
p
has
e Lear
nin
g
Acquir
es
Multi
tud
e of Gen
eralizat
ion Capabi
lities i
n
Controll
ing
Robot
Joint
without
Expon
ential Gr
owth in
Comple
xity
Lav
dim
Kur
t
aj, Vj
os
a Sh
at
ri
,
Ili
r Li
man
i
Facul
t
y
of Electr
ic
a
l
and
Com
put
er
Eng
ineeri
ng,
Univer
sit
y
of
Pri
shtina
“
Hasan
Pr
ishti
na”
,
Kos
ovo
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Ja
n
18
, 2
01
8
Re
vised
Ju
l
3
,
201
8
Accepte
d
J
ul
11
, 2
01
8
Proce
ss
ing
in
th
e
c
ere
b
el
lum
is
r
oughl
y
desc
r
ibed
as
fe
ed
forwar
d
proc
essin
g
of
inc
om
ing
inf
orm
at
ion
ov
er
t
hre
e
lay
e
rs
of
th
e
c
ere
b
el
l
ar
cor
t
ex
th
at
send
int
ermedi
at
e
ou
tput
to
dee
p
c
e
reb
ellar
nu
cl
e
i,
the
on
l
y
ou
tp
ut
from
the
ce
reb
el
lum
.
B
esi
de
thi
s
m
ai
n
p
ic
t
ure
the
r
e
ar
e
sev
era
l
f
ee
db
ac
k
ro
u
te
s,
m
ain
l
y
not
included
in
m
odel
s
.
In
th
is
p
ape
r
we
use
new
m
odel
for
neur
o
nal
ci
rcu
it
o
f
the
ce
r
ebella
r
gr
anul
e
c
el
l
l
a
y
er
,
as
co
ll
e
ct
ion
of
i
dea
l
iz
ed
gr
anul
e
c
el
l
–
g
o
lgi
cel
l
bu
il
ding
blo
cks
with
c
apa
b
il
i
t
y
of
gen
erati
ng
m
ult
i
-
dimensional
receptiv
e
fie
lds
m
odulated
b
y
sepa
rate
inpu
t
coming
to
lowe
r
dendr
i
te
tree
of
Golgi
ce
l
l
.
Result
ing
c
ere
b
el
l
ar
m
odel
controlle
r
with
two
-
phase
le
arn
ing
will
ac
qu
ire
m
ult
it
ude
of
g
e
ner
alizati
on
ca
p
abi
litie
s
when
u
sed
as
robot
jo
i
nt
cont
ro
ll
er
.
Thi
s
wi
ll
usual
l
y
req
u
ire
m
ore
th
a
n
one
Purkin
je
c
el
l
pe
r
output.
Functi
on
al
i
t
y
of
gra
nule
cell
-
G
olgi
c
el
l
bu
il
ding
bloc
k
was
ev
al
u
at
ed
wi
th
sim
ulations
using
Sim
uli
nk
single
compart
m
ent
s
piki
ng
n
eur
ona
l
m
odel
.
Tr
ai
n
e
d
av
era
ging
ce
reb
el
l
ar
m
odel
cont
ro
ller
at
t
ai
n
s
ver
y
good
tr
acking
resul
ts
for
wide
r
ange
of
unlearne
d
slo
wer
and
faste
r
tr
aj
e
ct
ori
es,
wi
th
addi
ti
on
al
improvem
ent
s
b
y
rel
e
arn
ing
a
t
f
aste
r
traje
ct
ori
es.
Inc
lusion
of
n
e
w
d
y
n
amical
ef
fec
ts
to
th
e
cont
roller
r
esult
s
with
li
n
ea
r
gro
wth
in
complex
i
t
y
for
inpu
ts
ta
r
get
ing
lower
dendr
ite
tr
ee
of
Golgi
c
el
l
,
important
for
cont
rol
appl
i
ca
t
ions
in
r
oboti
cs,
but
not
onl
y
.
Ke
yw
or
d:
Av
e
ra
ging cere
bella
r
m
od
el
Ce
reb
el
la
r g
ra
nu
le
cell
lay
er
Ce
reb
el
la
r
m
od
el
co
ntr
oller
Gen
e
rali
zat
ion
Go
l
gi cell
lo
w
er
dendr
it
e tree
Gr
a
nule
cell
-
G
olg
i cel
l ci
rc
uit
Modula
te
d rec
eptive
fiel
ds
Pu
r
ki
nj
e
cell
Robot
j
oi
nt contr
oller
Sing
le
c
om
par
tm
ent
m
od
el
Tw
o
-
ph
a
se lea
rn
i
ng
Copyright
©
201
8
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Vjos
a
Shat
ri,
Faculty
of Elec
tric
al
an
d C
om
pu
te
r
E
ng
i
neeri
ng
,
Un
i
ver
sit
y o
f P
rishtina “
Hasa
n
P
rishtina”
,
Breg
u
i
Diel
li
t p.n., 1
0000
Pr
i
sh
ti
na, K
osovo
.
Em
a
il
:
vj
os
a
.s
hatri@
un
i
-
pr.e
du
1.
INTROD
U
CTION
Ar
ti
fici
al
ne
ural
netw
orks
a
re
a
n
im
po
rtant
bra
nch
of
inf
or
m
at
ion
pr
ocessin
g.
It
st
arted
wit
h
insp
irat
io
n
f
rom
l
iving
ne
uro
nal
ci
rc
uits
pr
e
sent
in
li
ving
or
gan
ism
s,
as
m
eans
f
or
so
l
ving
pro
blem
s
of
te
c
hn
ic
al
i
m
po
rtance
th
at
seem
to
b
e
ha
ndle
d
so
e
asi
ly
by
them
.
Attract
ive
wa
s
the
way
of
so
lvi
ng
pr
ob
le
m
s
by
m
i
m
ic
kin
g
t
he
le
arn
in
g
proce
s
s,
without
e
xpl
ic
it
l
y
deali
ng w
it
h
t
he
c
om
plexity
of
t
he
prob
le
m
.
Des
pite
init
ia
l
enth
u
sia
sm
fo
ll
ow
e
d
by
disap
po
i
ntm
ent
and
a
pe
rio
d
of
al
m
os
t
aban
do
ne
d
a
rtific
ia
l
ne
ural
net
wor
k
res
earch
,
it
r
evive
d
a
nd
dev
el
op
e
d
t
o p
ow
e
rful to
ol
for
s
olv
in
g wide
range
of tec
hni
cal
p
r
oble
m
s.
Ce
reb
el
lum
play
ed
an
im
po
r
ta
nt
ro
le
in
t
his
process
.
It
ga
ve
ho
pe
to
de
ci
ph
e
r
inter
na
l
fu
nc
ti
on
al
pr
i
nciples
of
ne
uro
nal
ci
rcu
it
s
[
1],
ba
sed
on
regular
an
d
si
m
ple
layered
a
rr
a
ng
em
ent
of
neur
on
s
acr
os
s
w
ho
le
cerebell
ar
co
rtex,
with
well
de
fine
d
in
puts
a
nd
outp
uts,
see
m
ing
ly
as
unif
or
m
rep
et
it
ion
of
sam
e
el
e
m
e
ntary
cerebe
ll
ar
neur
on
al
ci
rc
uit
bet
ween
gro
up
s
of
ne
uro
ns
.
Af
te
r
init
ia
l
inf
orm
a
ti
on
a
bout
a
natom
y
and
physi
ology
,
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N:
20
88
-
8708
Ce
rebell
ar
M
odel
Co
ntro
ll
er
wi
th n
ew
M
ode
l of Gra
nule
...
(
La
v
dim
K
ur
t
aj
)
4293
two
the
or
ie
s
of
ce
reb
el
la
r
le
arn
i
ng
em
erg
e
d,
at
1969
by
Ma
rr
[
2]
a
nd
at
1971
by
Al
bu
s
[3
]
.
Bot
h
m
od
el
s
assum
ed
existe
nce
of
s
pecific
physi
ol
og
ic
al
chan
ges
at
s
pe
ci
fic
sit
e
du
rin
g
le
a
rn
i
ng,
fac
t
disc
ov
e
red
for
t
he
first
tim
e
at
1982
b
y
It
o
[
4],
a
nd
h
a
pp
e
ni
ng
i
n
the f
orm
of
long
-
te
rm
dep
r
ession
(LT
D), as
pr
e
dicte
d
by
Albus
[3
]
.
Be
cause
o
f
these
c
on
t
rib
ut
ion
s,
s
om
et
i
mes
it
is
re
ferred
to
as
the
Ma
r
r
-
Al
bu
s
-
I
t
o
t
he
or
y
[
5].
Lat
e
r
a
t
1975
Albus
f
or
m
al
ized
t
his
t
heory
with
c
om
pu
ta
ti
on
al
m
od
el
,
known
as
Ce
reb
e
ll
ar
Mo
del
Ar
t
ic
ulati
on
C
ontr
oller
[6
]
,
with
la
st
at
tribu
te
s
ori
gin
at
in
g
from
assum
ed
m
ain
functi
onal
i
nvolv
em
ent
of
the
ce
rebel
lu
m
in
coor
din
at
in
g
c
om
plex
m
ulti
-
j
oi
nt
m
ov
em
e
nts.
This
creat
ed
ne
w
br
a
nc
h
i
n
the
group
of
arti
fici
al
ne
ur
a
l
netw
orks, the
CM
AC n
e
ur
al
netw
ork, o
r
si
m
pl
y C
MAC. Since init
ia
l i
nt
rod
uction, it
was at
tract
ive fo
r
it
s f
as
t
le
arn
in
g
a
nd
possibil
it
ie
s
for
si
m
ple
real
-
tim
e
i
m
ple
m
entat
i
on
s
.
R
oute
of
dev
el
op
m
ent,
even
t
hought
i
niti
al
ly
to
be
bio
lo
gica
ll
y
based
,
f
ollo
wed
with
m
any
var
ia
ti
ons
not
necessa
rily
sti
c
king
to
bio
lo
gi
cal
plausib
il
it
y.
Most
of
CM
AC
a
pp
l
ic
at
ion
s
a
re
co
ntr
ol
base
d,
as
init
ia
l
functi
on
al
assum
ption
for
c
ere
bellum
,
bu
t
they
s
pan
wi
de
range o
f
ap
plic
at
ion
s [
7],
sim
i
la
r
to n
e
w
fi
nding
s
f
or
fun
ct
i
on
al
i
nvolv
em
ent
of
ce
rebel
lum
to
oth
e
r
f
unct
ions,
beyo
nd stere
otyped o
ne,
up t
o
c
ogniti
on a
nd em
otion
[8
]
.
CM
AC
belo
ng
s
to
a
fam
il
y
of
neural
net
wor
ks
with
lo
cal
le
arn
i
ng.
All
of
t
hem
su
ff
e
r
from
expo
nen
ti
al
increase
in
c
om
plexit
y
wh
en
dim
ension
al
it
y
of
the
in
put
sp
ace
inc
rease
s.
They
pa
rtit
ion
m
ulti
-
dim
e
ns
io
nal
input
s
paces
i
nto
m
ulti
-
dim
e
ns
io
nal
par
ti
ti
on
s
,
e
.g.
hype
rcubes
or
hype
r
-
s
pheres
,
whose
num
ber
in
creases
expo
nen
ti
al
ly
wh
il
e
i
ncr
easi
ng
num
ber
of
in
pu
t
sp
ac
e
dim
e
ns
io
ns
.
T
his
w
as
note
d
si
nce
i
niti
al
m
od
el
proposal
and
so
m
e
so
lut
ion
wer
e
gi
ven.
S
om
e
of
the
m
red
uce
re
pr
e
sentat
ion
capa
bili
ti
es
of
the
ne
tworks
,
wh
il
e
oth
e
r
s
introd
uce
po
s
s
ibil
it
y
of
inter
f
eren
ce
bet
wee
n
nons
im
il
ar
l
e
arn
i
ng
that
is
m
anifested
as
noise
.
E
ve
n
a
r
ough
par
ti
ti
on
i
ng
of
m
ul
ti
-
di
m
ensio
nal
in
pu
t
sp
ace
m
a
y
fast
gro
w
to
im
pr
act
ical
nu
m
ber
of
pa
rtit
ion
s.
H
ow
br
ai
n
so
lves
this
pro
blem
wh
il
e
it
ha
nd
le
s
m
uch
hi
gh
e
r
dim
ension
s
t
ha
n
we
us
e
f
or
sim
il
ar
te
chn
ic
al
pro
ble
m
s,
li
ke
in
rob
otics
?
Si
nce
struct
ur
al
c
on
te
nt
of
whol
e
brai
n
is
not
unif
or
m
,
but
with
m
any
va
riat
ion
s
to
dif
fer
e
nt
pa
rts
with
pres
um
ed
diff
e
re
nt
func
ti
on
,
br
i
ng
s
t
he
pro
blem
to
r
el
at
ion
bet
wee
n
str
uctu
re
an
d
functi
on,
a
nd
wh
ic
h
determ
ines
w
hi
ch
[
9].
Ultra
str
uctu
ral
a
nd
m
olecular
dif
fer
e
nc
es
in
t
he
s
eem
ly
si
m
il
ar
br
ai
n
ci
rc
uits,
sp
eci
f
ic
al
ly
to
the
cere
bellum
[1
0],
m
ay
po
i
nt
to
so
m
e
m
or
e
custom
ized
ci
rc
uits
for
sp
eci
fic
functi
on
s
that
w
ou
l
d
ena
bl
e
m
or
e
op
ti
m
a
l
so
luti
ons
for
s
pecific
pr
ob
le
m
s.
In
this
co
ntex
t
a
pp
li
cabi
li
ty
of
le
arn
e
d
inf
or
m
at
ion
to
si
m
il
ar
sit
uations
,
i.e.
gen
e
rali
zat
ion
[11], is
highly
desira
ble pr
op
e
rty
.
Wh
at
is
f
un
ct
i
on
al
sig
nifica
nc
e of
gra
nu
le
c
el
l l
ay
er for
ge
ner
al
iz
at
ion a
bi
li
t
ie
s of t
he
ce
reb
el
lum
can
be
li
nk
e
d
with
at
tribu
te
d
f
un
c
ti
on
of
this
la
y
er
withi
n
di
ff
e
r
ent
(stat
ic
)
cer
ebell
ar
m
od
el
s.
Ge
ner
al
pictu
re
f
or
m
ajo
rity
of
ce
r
ebell
ar
m
od
el
s
is
that
al
l
in
form
ation
e
nteri
ng
the
cere
bellum
through
m
os
sy
fibers
will
ta
rg
et
gr
a
nule
cel
ls
a
nd
Go
l
gi
cel
ls,
resi
ding
in
th
is
la
ye
r,
an
d
de
ep
c
ere
bella
r
nu
cl
ei
[
12
]
,
[
13
]
,
[
14
]
,
[
15
]
.
Thi
s
incom
ing
in
for
m
at
ion
is
com
bin
at
ori
al
ly
ex
pande
d
[2
]
,
[
3]
for
se
ve
ral
or
der
s
of
m
agn
it
ud
e
at
outp
ut
of
t
his
la
ye
r,
bein
g
in
form
ation
car
ried
by
pa
rall
el
fibers.
Mo
dels
assum
e
that
thi
s
is
the
place
wh
e
re
act
i
vity
sp
ars
e
higher
-
or
der
re
pr
ese
ntati
ons a
re
gen
e
rated
fr
om
low
er
-
orde
r
re
presentat
io
ns
on m
os
sy fibers.
Sp
a
rsen
e
ss
[16
]
is
co
ns
ide
red
as
m
eans
f
or
optim
iz
ing
sto
ra
ge
ca
pacit
y
of
fo
ll
owin
g
netw
ork,
an
d
as
m
eans
to
m
ake
le
arn
i
ng
easi
e
r.
Othe
r
a
sp
ect
of
t
hi
s
la
ye
r
i
s
spa
rse
co
nne
ct
ivit
y
pr
ese
nt
in
gr
a
nu
le
cel
ls,
the
m
os
t
nu
m
ero
us
ne
uro
ns
in
t
he
brai
n,
hav
i
ng
in
m
ean
f
our
s
hort
dend
rite
s
for
receivi
ng
inf
or
m
at
ion
f
r
om
fo
ur
disti
nct
m
os
sy
fibers.
It
was
s
how
n
i
n
[
17
]
t
hat
for
fee
dforwar
d
net
works
sim
i
la
r
to
ce
re
bella
r
processi
ng
this
is
require
d
f
or
eff
ic
ie
nt
decorr
el
at
ion
an
d
patte
rn
se
pa
rati
on.
Ma
ny
cerebell
ar
m
od
el
s
fo
ll
owin
g
CM
AC
a
ssu
m
e
that
in
form
ation
ca
rr
ie
d
by
m
os
sy
fiber
s
is
po
pu
la
ti
on
c
od
e
d,
wh
e
re
e
ach
m
os
sy
fi
be
r
has
rece
ptiv
e
fiel
d
descr
i
bed
with
so
m
e
one
-
dim
ensio
nal
basis
functi
on
(e.
g.
s
q
ua
re,
tria
ng
ul
ar
a
nd
Ga
us
sia
n).
P
r
ocessin
g
each
com
bin
at
ion
of
m
os
sy
fiber
s
inputs
with
m
ulti
plica
ti
on
operat
or
is
ty
pical
fo
r
m
od
el
s
of
gr
a
nu
le
cel
l
la
ye
r,
resu
lt
in
g
in
ou
tpu
t
of
c
orres
pondin
g
num
ber
of
m
ulti
-
di
m
e
ns
io
nal
rece
pt
ive
fiel
ds
at
pa
rall
el
fiber
s
,
m
akin
g
gen
e
rali
zat
ion
local
.
This
is
t
he
place
w
her
e
sp
ar
se
co
nnec
ti
vity
,
pr
ese
nt
in
bi
ologica
l
eq
uiv
al
ent,
will
he
lp
t
o
pr
e
ve
nt
com
bin
at
ori
al
ex
plo
s
ion
.
E
xtension
s
that
c
hange
c
onve
ntion
al
C
MAC
str
uctu
re
,
but
k
eep
fee
dfo
rw
a
r
d
processi
ng,
m
od
i
fy
the
way
the
ou
t
pu
ts
of
m
ulti
-
di
m
ension
al
receptive
fiel
ds
a
re
pro
cesses.
F
r
om
s
i
m
ple
m
ul
ti
plica
ti
on
with
c
orres
pondin
g
weig
ht,
it
is
proce
ssed
w
it
h
m
ulti
-
var
ia
ble
first
-
degree
poly
nom
ial
[18],
or
si
m
il
ar
[1
9
]
,
[
20]
.
Co
ns
t
a
nt
m
e
m
ber
of
poly
nom
ial
pr
ese
rv
e
s
f
un
ct
io
nalit
y
of
co
nventio
na
l
CM
AC.
Fi
rst
orde
r
te
rm
s
intro
duc
e
the
need
for
ad
diti
onal
in
puts,
bei
ng
rate
co
de
d
ve
rsion
s
of
in
form
at
i
on
that
tra
ns
f
orm
s
to
popula
ti
on
c
od
e
befo
re
ente
ring
t
he
ce
reb
el
l
um
as
m
os
sy
fiber
s
.
T
hese
ne
w
in
pu
ts
will
weig
ht
li
near
ly
ou
t
pu
ts
of
t
he
gran
ule
cel
l l
ay
er.
Pr
es
ence of m
os
sy fiber
s w
it
h
t
wo ty
pes
of
c
odin
g,
popula
ti
on
c
od
e
d
a
nd
rate c
od
e
d,
seen
bi
ologica
ll
y
poses
no
pr
ob
le
m
,
since
si
ng
le
popula
ti
on
c
ode
d
m
os
sy
fi
ber
us
es
rate
co
de
to
repr
es
ent
it
s
act
ivit
y.
Moss
y
fiber
s
ta
r
geting
dir
ect
ly
locat
ion
s
beyo
nd
gr
a
nule
cel
l
la
ye
r
is
not
bio
lo
gical
ly
plausibl
e.
Ne
w
m
od
el
of
pro
c
essing
i
n
the
gran
ule
cel
l
la
yer
gi
ves
pa
rtic
ular
im
po
rtanc
e
to
gran
ule
ce
ll
-
Go
l
gi
cel
l
neu
r
onal
ci
rcu
it
[
21]
,
as
basic
buil
ding
blo
c
k
of
gr
a
nule
cel
l
la
ye
r.
Go
l
gi
cel
l,
as
central
act
or
i
n
this
ci
rc
uit,
re
cei
ves
excit
at
or
y
fee
dback
in
puts
fro
m
par
al
le
l
fibers
(t
hro
ugh
up
pe
r
de
ndrite
tree
)
a
nd
e
xcita
tor
y
fee
dforwar
d
inputs
from
m
os
sy
fib
ers
(th
rou
gh
lo
wer
de
ndrite
t
r
ee).
La
te
r
is
ty
pical
ly
consi
de
red
to
be
from
the
s
am
e
set
of
m
os
sy
fibers
that
ta
r
ge
t
gr
a
nu
le
cel
l
t
hat
will
be
i
nh
i
bited
f
r
om
sa
m
e
G
olg
i
cel
l.
N
ew
m
od
el
assu
m
es
that
m
os
sy
fibe
rs
ta
rg
et
in
g l
ow
e
r
de
ndrite
tree
i
s
f
unct
ion
al
ly
disti
nct
in
form
at
ion
from
inform
atio
n ca
rr
ie
d
by
m
os
sy
fibe
rs t
hat
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
20
88
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
201
8
:
4292
-
4309
4294
ta
rg
et
c
orres
pond
i
ng
gra
nu
le
cel
ls.
More
ove
r
c
od
i
ng
in
t
wo
set
s
of
m
os
sy
fibers
is
diff
e
r
ent,
bein
g
rate
cod
e
d
for
t
hose
ta
r
get
ing
lo
wer dend
rite
tree.
F
un
ct
ion
al
ly
this
in
put wil
l
ser
ve
t
o
co
ntr
ol
m
ulti
plica
ti
vely
act
ivit
y
o
f
gr
a
nule
cel
l
outp
uts
form
ing
ci
rc
uit
with
c
orres
pondin
g
Gr
a
nule
cel
l.
Existence
of
t
hese
seg
re
gated
a
nd
functi
onal
ly
disti
nct
set
s
of
m
os
sy
fiber
s
was
la
te
r
sug
ge
ste
d
in
[
22]
as
exp
la
natio
n
f
or
in
vivo
a
nd
in
vitr
o
exp
e
rim
ents
they
per
f
orm
ed.
They
al
so
n
a
m
e
fu
nctio
n
of
seg
regat
ed
c
onnecti
vity
as
“par
al
le
l
fee
dforwar
d
inh
ibit
io
n”,
c
ontrast
ed
t
o
f
unct
ion
of
ste
re
otyped
c
onnecti
vi
ty
as
“cl
assic
a
l
feed
f
orwa
rd
i
nh
i
biti
on
”.
Thi
s
ne
w
m
od
el
pu
ts
po
l
ynom
ia
l
pr
oce
s
sing
of
gra
nu
le
cel
l
la
ye
r
as
bi
ologica
ll
y
plau
sibl
e
a
nd
perform
ed
inside
the
sam
e
la
ye
r
[23].
It m
ay
also contri
bute
to
s
pa
rsen
e
ss of acti
vity
issue at this
lay
er th
at
was rece
ntly
ch
al
le
ng
e
d [24
]
,
[
25]
.
In
this
pa
per
gran
ule
cel
l
la
ye
r
is
c
onside
re
d
as
colle
ct
io
n
of
gran
ule
cel
l
-
G
olg
i
cel
l
bu
i
lding
bl
oc
k
s.
C
apab
il
it
y
of
t
his
bl
ock
i
n
ge
ner
at
in
g
m
ulti
-
dim
ension
al
re
cepti
ve
fiel
ds
m
od
ulate
d
by
separ
at
e
i
nput
com
ing
to
lo
we
r
de
ndri
te
tree
of
Go
l
gi
cel
l
was
evalu
at
ed
with
sim
ulati
on
s
usi
ng
Si
m
ul
ink
sin
gle
c
om
par
t
m
ent
sp
iking
neur
on
al
m
od
e
l.
Ge
ner
al
iz
at
ion
cap
a
bili
ti
es o
f r
es
ulti
ng
a
ve
rag
i
ng
ce
re
bella
r
m
od
el
co
nt
r
oller w
it
h t
w
o
-
ph
a
se
le
arn
in
g wil
l
be
explo
red o
n
a
robot
j
oi
nt p
la
nt in
c
om
pu
te
d t
orq
ue
c
onfig
urat
ion.
2.
RESEA
R
CH MET
HO
D
Ce
reb
el
lum
,
or
li
tt
le
br
ai
n,
is
p
rese
nt
in
al
l
ve
rtebr
at
es
an
d
i
n
m
os
t
cases w
ei
gh
ts
a
bout 10
pe
rcen
t of
total
brai
n
w
e
igh
t.
O
n
the
oth
e
r
si
de
it
c
on
ta
in
s
m
or
e
than
hal
f
of
t
he
t
otal
nu
m
ber
of
ne
urons
in
the
br
ai
n
[
26]
.
F
unct
ion
al
ly
it
is
associat
ed
a
nd
is
co
ns
ide
r
ed
as
par
t
of
m
oto
r
structu
r
e.
It
is
im
po
rtant
in
m
ai
ntaining
postur
e
a
nd
balance,
i
n
fl
ue
nt
execu
ti
on
of
vo
lun
ta
ry
a
rtic
ula
te
d
m
ov
em
ent
s,
in
m
oto
r
le
arn
i
ng
,
and
base
d
on
c
onnecti
ons
it
m
akes
with
oth
e
r
par
ts
of
the
brai
n
it
m
ay
be
i
nvolv
e
d
in
m
any
oth
e
r
f
un
ct
i
on
s
,
up
to the c
ogniti
ve
level [
27
]
,
[
8]
.
Sp
iki
ng
m
od
el
of
ce
reb
el
la
r
gr
a
nule
cel
l
la
ye
r
f
or
Sim
uli
nk
wa
s
dev
el
oped
to
assess
the
a
bili
ty
of
gr
a
nule
cel
l
-
G
olg
i
cel
l
bu
il
di
ng
bl
ock
in
gen
e
rati
ng
m
od
ulate
d
higher
-
orde
r
rece
ptive
fiel
ds
.
A
ve
rag
i
ng
functi
onal
eq
uiv
al
ent
is
us
e
d
t
o
buil
d
new
a
ve
rag
i
ng
ce
reb
el
la
r
co
ntro
ll
er
[
28
]
,
[
23
]
.
Ro
bo
t
plant
[
28
]
f
or
te
sti
ng
global ge
ne
rali
zat
ion
pro
per
ty
to jo
i
nt s
peed
and joint a
ccel
erati
on is prese
nted.
2
.
1.
Neur
onal
Circui
t of t
he
Cerebell
um
In
c
ontrast
to
num
ber
of
ne
uron
s
co
ntaine
d
i
n
the
ce
re
bellu
m
[2
6],
num
ber
of
neuron
ty
pe
s
is
sm
a
ll
,
com
m
on
to
al
l
ver
te
br
at
es.
It
ho
sts
tw
o
e
xtre
m
e
ty
pes
of
ne
uro
ns
by
siz
e,
Pu
r
ki
nj
e
cel
l
a
nd
gra
nu
le
cel
l
.
First
is
known
a
s
la
r
gest
ne
uro
n
in
the
brai
n,
w
hi
le
seco
nd
one
a
s
sm
allest
and
m
os
t
nu
m
ero
us
ty
pe.
All
of
t
hem
are
arr
a
ng
e
d
in
highly
or
gan
iz
e
d
t
hr
ee
la
ye
r
str
uc
ture,
nam
ed
cer
ebell
ar
c
ort
ex.
Mi
dd
le
la
ye
r
one
-
cel
l
t
hick
c
onta
ins
on
ly
P
urkin
j
e
c
el
ls.
Nex
t
la
ye
r
in
the
i
nn
e
r
sid
e
m
a
inly
inh
ab
it
s
gr
a
nu
le
ce
ll
s
an
d
G
olg
i
cel
ls.
Thir
d,
oute
r
la
ye
r
of
cere
bella
r
c
or
te
x
is
know
n
as
m
olecular
l
ay
er.
Ty
pical
f
or
this
la
ye
r
is
pr
ese
nce
of
huge
nu
m
ber
of
pa
rall
el
fibers
that
cr
oss
at
rig
ht
an
gle
thr
ough
flat
s
hap
e
d
de
ndrite
trees
of
P
urki
nj
e
cel
ls.
Paral
le
l
fiber
s
a
re
i
n
fact
gr
a
nule
cel
l
ax
on
s
that
rise
ve
rtic
al
ly
fr
om
gran
ule
cel
l
la
ye
r,
t
hroug
h
Pur
kinje
cel
l
la
ye
r
,
an
d
w
he
n
r
ea
chin
g
the
m
olecular
la
ye
r
bif
ur
cat
e
i
n
T
-
s
ha
pe
d
f
orm
.
This
descr
ipti
on
of
gran
ul
e
cel
l
axo
n
is
from
final
m
at
ur
e
sta
te
po
i
nt
of
vie
w.
Bi
olo
gical
ly
route
is
opposit
e
[29],
f
ro
m
neuroge
nesis
at
th
e
tem
po
rar
y
e
xt
ern
al
gran
ular
la
ye
r,
thr
ough
m
igrati
on
on
m
olecular
la
ye
r,
t
o
fin
al
sta
te
in
the
gran
ular
la
ye
r
.
All
ci
rcu
it
for
m
at
ion
proce
ss
is
ve
ry
well
or
c
hestrat
ed
in
ti
m
e
and
sp
ace,
resu
lt
in
g
in
this
s
ophis
ti
cat
ed
functi
onal
struct
ur
e
,
init
ia
ll
y
tho
ug
ht
si
m
ple
to
disce
rn,
but
it
con
ti
nues
to
sur
pr
ise
sci
entifi
c
com
m
un
it
y
with
fi
ne
r
str
uctu
ral
de
ta
il
s
and
m
olecular
div
e
rsiti
es [
10]
. P
r
esent i
n
m
olecular lay
er a
r
e also
bas
ket c
el
ls and ste
ll
at
e cel
l
s.
Si
m
ple sk
et
ch i
n
Fig
ur
e
1 (m
od
i
fied fr
om
[
21
])
gi
ves
c
onne
ct
ion
s
betwee
n
cere
bella
r ne
uro
ns
. T
here
are
on
ly
t
wo
r
ou
te
s
w
he
re
i
nfo
rm
ation
ca
n
enter
c
ere
bellu
m
,
m
os
sy
fibers
an
d
cl
i
m
bin
g
fibe
rs.
Mo
ssy
fibers
form
m
any
con
ta
ct
s
with
gran
ule
cel
ls,
G
olg
i
c
el
ls,
an
d dee
p ce
re
bella
r
nucl
ei
. C
lim
bin
g fiber
sp
li
ts
in
s
evera
l
br
a
nc
hes,
a
nd
each
br
a
nc
h
m
akes
ver
y
strong
co
nnect
ion
(co
m
po
se
d
of
hundre
ds
of
c
onta
ct
s)
with
P
urki
nj
e
cel
l,
bu
t
e
ver
y
Purk
i
nj
e
cel
l
in
m
at
ur
e
sta
te
m
akes
exclusive
c
onnecti
on
with
on
ly
on
e
cl
i
m
bin
g
fi
ber
.
All
cl
i
m
bin
g
fibe
r
s
ori
gi
nate
from
infer
io
r
ol
ivary
nucl
eus
resi
ding
outs
ide
of
the
cer
ebell
um
,
bu
t
that
is
reciprocal
ly
co
nn
ect
e
d
with
de
ep
ce
reb
el
la
r
nu
cl
ei
.
C
onta
ct
s
betw
een
m
os
sy
fibe
rs,
gr
a
nule
cel
l
de
ndrite
s,
a
nd
Go
l
gi cell
axo
ns
a
r
e
done wit
hin
glo
m
eru
li
.
So
le
outp
ut
f
r
om
the
cerebell
ar
cor
te
x
is
th
rou
gh
Purk
i
nj
e
cel
l
axons.
M
os
t
of
them
,
after
sen
di
ng
colla
te
rals
bac
k
to
ce
re
bella
r
cor
te
x,
te
rm
inate
to
dee
p
ce
re
bella
r
nucl
ei
,
a
nd
f
ro
m
them
t
o
oth
er
pa
rts
of
the
br
ai
n.
Re
m
ai
nin
g
pa
rt
of
the
m
m
ake
si
m
ilar
neuronal
ci
r
cuits
with
ve
s
ti
bu
la
r
nucl
ei
.
Groups
of
ce
r
ebell
ar
neur
on
s
with
m
ic
ro
ci
rcu
it
ry
as
in
Fig
ure
1
a
re
orga
nize
d
at
higher
le
vel
in
m
ic
ro
zo
nes
[
30]
,
an
d
e
ven
hi
gh
e
r
t
o
zon
e
s,
that
a
re
thought
to
ha
ve
sp
eci
fic
f
unc
ti
on
s
dep
e
nd
in
g
on
t
he
in
pu
t
and
outp
ut
co
nnect
ions
they
r
ecei
ve
from
,
and
sen
d
to o
the
r
par
ts
o
f
the n
e
rvo
us
syst
e
m
.
Ther
e
are
ver
y
s
pecif
ic
at
tribu
te
s of co
nnect
io
ns
b
e
tween
act
or
s
of
the
ce
reb
el
lum
. Fi
rst,
ve
ry
high
di
ve
rg
e
nce
of
m
os
sy
fibe
r
in
pu
t
in
form
ation
,
f
r
om
m
il
l
ion
s
of
m
os
sy
fibers
t
o
bill
io
ns
of
the
gra
nule
cel
ls
[
27
]
.
Seco
nd,
co
nve
rg
e
nce
f
ro
m
hund
red
s
of
th
ousa
nd
s
par
al
le
l
fi
ber
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N:
20
88
-
8708
Ce
rebell
ar
M
odel
Co
ntro
ll
er
wi
th n
ew
M
ode
l of Gra
nule
...
(
La
v
dim
K
ur
t
aj
)
4295
inf
or
m
at
ion
to
sing
le
outp
ut
f
ro
m
Purk
i
nj
e
c
el
l.
Furthe
r
c
onve
r
gen
ce
f
r
om
sever
al
hu
ndre
ds
of
Pur
kin
j
e
cel
ls
ou
t
pu
ts
t
o
one
deep
c
ere
bella
r
nu
cl
ei
.
Re
la
ti
ve
nu
m
ber
of
di
verge
nce
an
d
c
onve
rg
e
nce
is
pr
ese
r
ved
but
num
ber
s
are s
pecies
depend
e
nt. A
bove nu
m
ber
s a
re
for hig
hly de
velo
ped s
pecies, li
ke
hum
ans.
Figure
1.
Ne
uron
al
ci
rcu
it
of the ce
reb
el
lum
, n
eu
r
on
s
, c
onne
ct
ion
s,
and si
gn
al
ro
utes
(m
od
ifie
d f
ro
m
[
21
])
.
PC
: P
urkin
j
e c
el
l;
GC
: g
ranul
e cel
l;
GoC
: G
olg
i cel
l;
StC
:
ste
ll
at
e cel
l
;
BaC
:
bas
ket cel
l;
Gl
: glom
eru
li
;
aa
:
gr
a
nule
cell
as
cend
i
ng axo
n;
PF
: pa
rall
el
f
ib
ers;
ud
t
:
Go
l
gi cell
u
ppe
r den
dr
it
ic
tree;
ldt
:
Go
l
gi cell
lo
w
er
dend
riti
c tree;
mf1
,
m
f2
,
m
f3
:
m
os
sy fibe
rs;
CF
: cl
im
bin
g
f
iber;
P
C axon
:
Purk
i
nje cel
l a
xon;
DCN
:
dee
p
cerebell
ar
nucl
ei
;
IO
: i
nf
e
rio
r
o
li
ve;
ML
: m
olecular
lay
er;
PC
: P
urkin
j
e c
el
l l
ay
er;
GL
: gr
a
nule
cell
lay
er;
WM
:
wh
it
e m
at
te
r
Ultra
str
uctur
e
of
c
onnecti
ons
betwee
n
dif
fere
nt
act
or
s
of
th
e
ci
rc
uit
is
pa
rt
of
inten
sive
re
search
,
a
nd
cou
l
d
hav
e
im
portant
im
plications
for
poss
ible
at
ta
inable
functi
onal
va
riat
ion
s.
I
n
this
sense
m
os
sy
fi
ber
s
,
accor
ding
to
[
21
]
,
are
se
par
a
te
d
int
o
gro
up
s
th
at
can
ta
r
ge
t
dif
fer
e
ntial
ly
sp
eci
fic
str
uc
tures
an
d/or
s
pecific
locat
ion
s
on
t
hem
,
base
d
on
act
or
a
nd
locat
ion
sp
eci
fic
m
olecular
m
ark
ers
[10].
Set
m
f1
of
m
os
sy
fib
ers
is
com
m
on
one
t
hat
ta
r
get
gr
a
nule
cel
ls.
They
can
ta
r
get
G
olgi
cel
ls
as
well
without
c
o
ns
eq
uen
ce
s
for
ne
w
m
od
el
.
Stand
a
r
d
m
od
e
l
assum
es
both
ta
r
gets.
Sp
eci
f
ic
to
t
his
m
odel
is
set
m
f2
a
nd
m
f3
of
m
os
sy
fibe
rs
that
t
arg
et
lowe
r
de
ndrite
tree
of
G
olg
i
c
el
ls
and
dee
p
c
ereb
el
la
r
nu
cl
e
i
,
resp
ect
ively
,
bu
t
no
t
the
c
orrespo
nd
i
ng
gro
up
of
gr
a
nule
cel
ls. Sp
eci
fic f
un
ct
i
on
al
im
po
rta
nc
e
was
at
trib
uted
for
set
m
f2
m
os
sy
fiber
s
i
n
[
21
]
. Later,
au
thors o
f
[22]
acco
rd
i
ng
to
their
fi
nd
i
ng
s
sugg
est
t
hat
separ
at
e
set
s
of
m
os
sy
fiber
s
m
ay
ta
rg
et
neig
hbori
ng
gr
a
nu
le
cel
ls
and Golgi cel
ls
that are
also
fun
ct
io
nall
y dis
ti
nct.
2.2
.
Simul
ink
Spiking
Mode
l fo
r
Gr
an
ule
Cell
s and
Go
l
gi C
el
l
s
To
a
naly
ze
inf
or
m
at
ion
proc
essing
at
gran
ule
cel
l
la
ye
r,
sp
iki
ng
m
od
el
s
f
or
t
wo
m
os
t
wide
spre
a
d
inh
a
bitants
of
t
his
la
ye
r
was
dev
el
op
e
d
i
n
S
i
m
ulink
,
i.e.
f
or
gra
nu
le
cel
ls
and
G
o
lgi
cel
l
s.
A
dopte
d
m
od
el
is
sing
le
com
par
t
m
ent
m
od
ifie
d
integrate
-
an
d
-
f
ire
that
com
pu
t
es
m
e
m
br
ane
pote
ntial
f
r
om
[3
1],
with
par
a
m
et
ers
from
the sam
e
publica
ti
on
m
r
e
s
t
r
e
s
t
m
G
A
B
A
G
A
B
A
m
N
M
D
A
m
N
M
D
A
N
M
D
A
m
A
M
P
A
A
M
P
A
m
m
V
E
G
V
E
t
g
V
E
V
g
t
g
V
E
t
g
dt
dV
C
,
(
1)
This
is
a
first
order
diff
e
re
ntial
equ
at
io
n
th
at
acco
un
ts
f
or
chem
ic
al
synapses
(
for
AMP
A,
NM
D
A
,
and
GA
B
A
rec
epto
rs)
an
d
le
akin
g
(r
est
in
g
c
onduct
ivit
y
G
r
est
).
C
m
is
m
e
m
br
ane
ca
pacit
ance.
Nat
ur
e
of
eac
h
m
e
m
ber
in
(
1)
is
current
a
nd
r
epr
ese
nts
postsy
nap
ti
c
c
urre
nt
s
cause
d
from
i
on
s
pa
ssin
g
th
r
ough
c
orrespo
nd
i
ng
ion
c
ha
nn
el
s.
g
AMPA
,
g
NMDA
and
g
GABA
are
co
nductance
s,
a
nd
E
AMPA
,
E
NMDA
an
d
E
GABA
are
rev
e
rsing
po
te
ntial
s
for
AMP
A,
N
MDA
a
nd
GAB
A
recept
ors
,
corres
pondin
gly
.
E
resr
is
resti
ng
m
e
m
br
ane
po
te
ntial
.
Cond
ucta
nce
of
NM
DA
cha
nn
el
is
d
e
pe
ndent on m
e
m
br
ane
vo
lt
age
[3
2] an
d
is
giv
e
n
a
s
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
20
88
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
201
8
:
4292
-
4309
4296
/
]
[
1
1
2
,
Mg
e
V
g
m
V
m
N
M
D
A
(2
)
with
c
onsta
nts
eq
ual
to
α
=6
2V
-
1
,
[M
g
2+
]
=1.2m
M
and
β
=3.57
m
M.
Ion
cha
nnel
co
nd
uctances
are
only
tim
e
dep
e
ndant
,
li
ke
in
s
pik
e
r
es
po
ns
e m
od
el
[3
3], and a
re
giv
e
n as f
ollows
s
t
t
s
I
C
C
s
I
C
C
t
t
e
t
g
t
t
t
g
I
C
C
s
,
,
0
/
)
(
(3
)
wh
e
re,
t
s
de
not
es
sp
ike
a
rr
i
va
l
tim
e
a
t
pr
esy
nap
ti
c
sit
e.
IC
C
sta
nd
s
for
ty
pe
of
io
n
c
hannel
co
nducta
nc
e,
f
or
AMPA,
NMD
A,
or
GA
B
A.
Cond
uctances
are
m
od
el
ed
as
fall
i
ng
e
xpone
ntial
functi
ons
with
m
axi
m
u
m
valu
es
g
AMPA
(
t
s
),
g
NMDA
(
t
s
)
a
nd
g
GABA
(
t
s
),
an
d
with
t
i
m
e
con
sta
nts
τ
AMPA
,
τ
NMDA
and
τ
GABA
,
c
orre
sp
on
dingly
.
Fi
rst
tw
o
cond
uctances
will
res
ult
wit
h
e
xcita
tory
a
nd
t
he
la
st
one
with
i
nh
i
bitor
y
postsy
na
ptic
c
urren
t.
Data
use
d
for
par
am
et
ers
in
eq
uatio
ns
for
gr
a
nu
le
cel
l
an
d
G
olg
i
c
el
l
m
od
el
are
the
one
give
n
in
Ta
ble
1
a
nd
Table
3
of
[
31
]
.
A
utho
rs
of
th
e
pa
per
ha
ve
c
ollec
te
d
these
dat
a
f
ro
m
m
an
y
sources
,
as
m
entione
d
at
thei
r
Ta
ble
1.
Fig
ur
e
2
s
ho
ws Si
m
ulink
m
od
el
for
gr
a
nu
l
e cel
l.
Go
l
gi ce
ll
Si
m
ulink
m
od
el
is
sim
il
ar,
bu
t
it
h
as
tw
o s
et
s of
AMPA
rece
ptors,
for
uppe
r
a
nd
lo
wer
de
ndr
it
e
tree
synap
s
es,
la
ck
of
NMDA
rece
ptors,
and
prese
nc
e
of
le
a
k
current a
nd s
ponta
neous
o
sci
ll
at
ion
[31,
34
]
.
(a)
(b)
(c)
Figure
2.
S
pik
i
ng gra
nu
le
cell
Sim
ulink
m
od
el
.
U
pp
e
r
sim
ulati
on
d
ia
gr
am
m
od
el
s ch
an
ne
l cond
uctance
. At
le
ft, f
irst t
wo bl
ock
s
m
od
el
exc
it
at
or
y p
os
tsy
nap
ti
c c
urre
nts
, i_AMPA
an
d i
_N
MD
A, ge
na
rated
from
pr
esy
nap
ti
c
pu
l
se ar
riving at
port
1 (E
xc)
fro
m
m
os
sy fiber. Thir
d bloc
k
m
od
el
s
in
hib
it
or
y cur
ren
ts
i_
G
ABA
from
p
ulse at
port
2 (In
h)
.
F
ort
h
bl
ock m
od
el
s leakin
g
c
urre
nt i_leak
. M
odel
o
f
so
m
at
ic
acti
on
is at t
he
bo
tt
om
rig
ht
, wit
h
s
ubm
od
el
s for
i
nteg
rati
on
of cu
rr
e
nts, g
ener
at
i
on
of s
pi
ke,
a
nd g
e
ne
ra
ti
on
of
re
fr
act
ory
per
i
od
Mod
el
was
bu
il
t
with
obj
ect
s
that
will
al
lo
w
processi
ng
of
m
ulti
di
m
en
sion
al
i
nputs
s
ign
al
s.
Eac
h
channel
co
nduc
ta
nce
is
m
od
e
le
d
acc
ordin
g
t
o
uppe
r
sim
ula
ti
on
dia
gr
am
.
F
our
bl
ocks
at
le
ft
m
od
el
fou
r
s
ources
of
syna
ptic
cu
r
ren
ts
acc
ordi
ng
to
(
1).
Mo
de
l
of
so
m
at
ic
ac
ti
on
is
giv
e
n
a
t
bo
tt
om
rig
ht.
Encircle
d
a
re
three
Int
e
g
r
a
t
i
o
n
S
pi
ke
ge
n
e
ra
t
i
o
n
R
e
f
r
a
c
t
o
r
y
pe
r
i
o
d
C
ur
r
e
n
t
s
In
t
e
rna
l
S
t
a
t
e
(M
e
m
b
r
a
n
P
o
t
e
n
t
i
a
l
)
dt
i
C
V
m
m
1
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N:
20
88
-
8708
Ce
rebell
ar
M
odel
Co
ntro
ll
er
wi
th n
ew
M
ode
l of Gra
nule
...
(
La
v
dim
K
ur
t
aj
)
4297
su
bm
od
el
s:
f
or
integ
rati
on,
s
pi
ke
gen
e
rati
on
and
for
refract
or
y
tim
e.
Sp
i
ke
is
i
deali
zed
a
s
re
ct
a
ngular
pulse
of
un
it
ary
am
plitu
de
a
nd
with
width
(Sw)
eq
ual
to
0.1
m
s.
Gain
of
i
on
ch
ann
el
c
onduct
a
nce
is
scal
ed
a
ccordin
g
to
the
se
value
s
.
P
ropa
gatio
n
of
s
pik
e
s
i
n
t
he
axon
is
not
m
od
el
ed.
T
hese
m
od
el
s
are
e
ncas
ed
i
ns
ide
tw
o
ty
pes
of
ne
uro
n
sub
syst
e
m
b
locks
with
diff
e
re
nt
processi
ng
of
excit
at
or
y
sy
nap
ti
c
c
urren
ts
,
Fig
ur
e
3.
Fir
st
on
e
,
Gr
C
_n
s
,
is
m
od
el
of
a
sin
gle
ne
uro
n
with
n
syna
ptic
in
puts.
Sec
ond
on
e
,
G
rC_n
n,
is
m
od
el
f
or
n
ne
uro
ns.
Su
m
m
a
ti
on
of
sp
ikes
ta
r
getin
g
de
ndrite
tree
for
la
te
r
case
is
m
od
e
le
d
ou
t
of
ne
uro
ns
bl
ock.
N
um
ber
of
sy
napse
s
in
the
fi
rst
m
od
el
an
d
num
ber
of
ne
uro
ns
in
the
seco
nd
m
od
el
will
be
dete
rm
ined
by
dim
ension
a
li
ty
of
corres
pondin
g
input
sig
nal.
Go
l
gi
cel
l
will
us
e
m
od
el
f
or
si
ng
le
ne
uro
n
with
m
any
s
ynapses
re
su
lt
ing
to
Go
C
_ns
subsy
stem
m
od
el
.
It
will
hav
e
tw
o
e
xcita
tory
in
pu
t
s,
from
up
pe
r
de
ndrite
tree
(
udt)
and
l
ow
e
r
de
ndrite
tree (ldt
).
Inhi
bi
tory in
pu
ts t
o Go
l
gi cell
[35
]
are
no
t m
od
el
e
d.
Figure
3. Tw
o form
s o
f gr
a
nu
le
cell
su
bs
yst
e
m
b
locks wit
h diff
e
re
nt
proc
e
ssing o
f
e
xcita
tory c
urre
nts
i_AMP
A
a
nd i
_N
M
DA to i
_E
xc: (
a
) proces
sing t
hat
will
generate m
od
el
of sin
gle n
e
uro
n wit
h
m
any
napses, s
ubsyst
e
m
G
rC_ns;
(b) pr
ocessin
g
t
ha
t wil
l generat
e m
od
el
f
or m
any n
e
uro
ns
.
Nu
m
ber
o
f sy
na
pse
s n
the f
ir
st m
od
el
an
d n
um
ber
of
neur
on
s
in
t
he se
co
nd m
od
el
w
il
l be
determ
i
ned b
y
dim
ension
al
it
y o
f
i_
Ex
c
2.3
.
Simul
ink
Spiking
Mode
l fo
r
Gr
an
ule
Cell
–
G
olg
i
Ce
ll
C
ir
cuit
Sp
iki
ng
m
od
el
s
of
gra
nu
le
ce
ll
and
Go
l
gi
ce
ll
are
us
e
d
to
bu
il
d
Sim
ulink
sp
iki
ng
m
od
el
of
gran
ule
cel
l
–
Go
l
gi
cel
l
ci
rcu
it
s
hown
i
n
Fi
gure
4,
w
he
re
Gr
C
_nn
is
m
od
el
fo
r
n
gr
a
nu
le
cel
l
ne
uro
ns
,
G
oC_ns
is
m
od
e
l
for
sin
gle
G
ol
gi
cel
l
with
n
sy
napses.
S
umm
at
ion
blo
c
k
wit
h
f
our
i
nputs,
x1_L
1
to
x1
_L
4,
represe
nts
gran
ule
cel
l
den
dri
te
s
receivin
g
in
pu
ts
fr
om
fo
ur
gro
ups
of
po
pula
ti
on
co
ded
m
os
sy
fiber
s.
Lo
wer
m
os
sy
fibers
corres
pond
to
o
ne
rate
co
de
d
m
os
sy
fiber
v1
_R
that
co
nnec
ts
to
l
ow
e
r
de
ndrite
tree
ldt o
f
G
olg
i
cel
l.
Sy
nap
ti
c
stren
gth
s
a
re
a
ssu
m
ed
to
be
c
on
sta
nt,
with
nom
inal
value
[
31
]
set
at
co
rr
e
sp
on
ding
neur
on
m
od
el
.
Gain
bl
ocks
S_
MF
-
G
rC,
S
_M
F_
G
oC,
a
nd
S_Go
C
-
G
rC,
c
urren
tl
y
set
to
1,
re
prese
nt
relat
ive
synaptic
s
tren
gth
.
T
hey
c
an
be
us
e
d
t
o
e
xplo
r
e
the
range
an
d
e
ffec
t
of
var
i
able
sy
nap
ti
c
s
tren
gth
s.
If
syn
aptic
stre
ngths
of
diff
e
re
nt
in
puts
t
o
gr
a
nule
cel
l
a
r
e
di
ff
e
ren
t,
eac
h
i
nput
(
of
f
ou
r
of
them
in
Fi
gure
4)
m
ay
ha
ve
it
s
ow
n
ga
in
blo
c
k
S_
M
F
-
G
rCi,
and
by
rem
ov
i
ng
S_
M
F
-
Gr
C
blo
c
k
afte
r
sum
m
a
ti
on
bl
ock.
Syna
pt
ic
stre
ng
t
h
S
_GrC
-
G
oC
is
us
e
d
to
c
orrect
eff
ect
ive
syna
pt
ic
stren
gth
[
31]
betwee
n
gr
a
nule
cel
l
ax
on
a
nd
up
per
de
ndr
it
e
tree
of
Go
l
gi
cel
l
in
acc
ord
ance
with
num
ber
of
gr
a
nule
cel
ls
ta
rg
et
in
g
Go
l
gi
cel
l.
Parall
el
fiber
s
c
onta
in
processe
d
m
ulti
di
m
ension
al
re
c
eptive
fiel
ds
(in
case
of
tw
o
-
dim
ension
al
i
nput
s
pa
ce
they
will
be
tw
o
-
dim
ension
al
rec
eptive
fi
el
ds
,
sho
wn
la
te
r
i
n
Figure
5).
Ge
ne
rati
on o
f
in
for
m
at
ion
with
p
opulati
on
c
ode
d
rece
ptive
fiel
ds
that
dri
ve
m
os
sy
fib
ers
is
t
houg
ht
no
t t
o be
par
t
of the
cere
be
ll
um
an
d
co
rr
es
ponds t
o pr
e
-
cer
ebell
ar s
tr
uctu
r
es.
2.4
.
M
os
s
y
Fi
bers
an
d
Info
rmat
i
on
E
nt
er
ing
th
e
Cereb
el
lum
Ther
e
are
t
wo
routes
w
he
re
i
nfor
m
at
ion
can
enter
t
he
cere
bellum
,
m
os
sy
fibe
rs
a
nd
cl
i
m
bin
g
fibe
rs.
In
m
os
t
cerebe
ll
ar
m
od
el
s
la
te
r
one
a
re
ass
um
ed
to
car
ry
error
i
nfor
m
at
ion
.
In
it
ia
l
m
od
e
l
and
m
os
t
of
m
od
el
var
ia
ti
ons
f
ollo
wing
assum
e
that
i
nfor
m
at
ion
car
ried
by
m
os
sy
fibe
rs
is
popula
ti
on
c
ode
d
[
6,
18]
.
I
n
thi
s
case
un
i
qu
e
i
nput
st
at
es
are
re
pres
ented
wit
h
uniqu
e
act
ivit
y
in
a
subset
of
m
os
sy
fibe
rs.
Si
m
plest
on
e
us
e
d
is
i
n
form
of
sev
era
l
on
e
-
dim
ension
al
la
ye
rs
com
po
s
ed
from
no
nove
rlap
ping
r
ecepti
ve
fiel
ds,
with
their
act
ivit
y
descr
i
bed
by
r
ect
angular
basi
s
f
un
ct
io
ns
.
I
n
this
case
eac
h
l
ay
er
will
ha
ve
o
nly
on
e
of
t
he
m
ac
ti
ve
for
ce
rtai
n
range
of
in
pu
t
values
, whe
re
act
ive m
eans for outp
ut
bein
g eq
ual to o
ne
a
nd zer
o for i
na
ct
ive case.
[
i
_
N
M
D
A
]
[
i
_
A
M
P
A
]
[
i
_
N
M
D
A
]
[
i
_
A
M
P
A
]
G
r
C
_
ns
E
x
c
I
n
h
A
x
o
n
Uc
G
r
C
_
nn
E
x
c
I
n
h
A
x
o
n
Uc
[
i
_
E
x
c
]
[
i
_
E
x
c
]
(b)
(a)
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
20
88
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
201
8
:
4292
-
4309
4298
Figure
4.
Sim
ulati
on
m
od
el
of
infor
m
at
ion
processin
g
i
n
the
sin
gle
gran
ule
cel
l
la
ye
r
bu
il
di
ng
blo
c
k
of th
e
cerebell
um
w
it
h
s
pik
i
ng n
e
uron m
od
el
s,
with
in
f
or
m
a
ti
on
rou
te
M
F →
Gr
C
-
G
oC →
PF
This
f
orm
of
cod
i
ng
was
pro
po
s
ed
by
Al
bus
[6
]
a
nd
la
ye
rs
are
known
a
s
Albus
ov
e
rlay
s,
Fig
ur
e
5.
Lat
er
la
ye
rs
w
ere
joine
d
t
o
a
sin
gle
la
ye
r
of
overlap
pi
ng
r
ecepti
ve
fiel
ds,
with
e
xtensi
on
t
o
m
any
s
ha
pes
of
basis
functi
ons
desc
ribi
ng
re
c
eptive
fiel
d
act
ivit
y.
A
no
t
her
view
is
thi
nk
i
ng
of
eac
h
rece
pt
ive
fiel
d
as
on
e
la
ye
r.
Re
gardless
of
how we
view
t
hem
, each
sig
na
l (u
s
ually
one
-
dim
ension
,
bu
t no
t
necessa
rily
)
will
b
e c
ode
d wit
h
a
gro
up
of
m
os
sy
fibe
rs
wh
e
re
s
ub
set
of
th
e
m
(o
ne
or
m
or
e)
will
be
m
or
e
or
le
ss
act
iv
e
in
a
s
pecific
rang
e
(ty
pical
ly
fo
rm
0
to
1)
.
M
os
s
y
fiber
s
ca
rr
yi
ng
rate
co
de
d
inf
or
m
at
ion
is
com
m
on
to
som
e
la
te
r
ver
sio
ns
of
cerebell
ar
[1
9,
20
]
or
hybri
diz
ed fuz
zy
-
cere
be
ll
ar m
od
el
s [1
8].
Gen
e
rati
on
of po
pu
la
ti
on c
od
e
d
sp
i
king
m
os
sy
fibers
f
r
om
inp
ut
sig
nal
is
done
with
tw
o
pro
cessi
ng
sta
ge
s,
uppe
r
par
t
i
n
Fi
gure
6.
First
sta
ge
from
inp
ut
s
ign
a
l
will
ge
ner
at
e
a
num
ber
of
ne
w
si
gn
al
s
with
sel
ect
ed
basis
f
unct
ion,
w
he
re
s
quare
,
tria
ngular
a
nd
Ga
us
sia
n
bein
g
m
os
t co
m
m
on
.
E
ach
sig
nal
is
then
co
nverte
d
to
st
ream
of
pulse
s
(sp
i
kes
)
at
seco
nd
sta
ge
,
w
he
re
pulse
f
reque
ncy
is
pro
portion
al
t
o
a
m
plit
ud
e
of
i
nput
sig
nal.
Ra
te
cod
e
d
s
pik
i
ng
m
os
sy
fiber
s
are
ge
ne
rated
by
sk
i
pp
i
ng
t
he
first
sta
ge,
i.e.
by
di
rectl
y
feed
i
ng
in
pu
t
sig
nal
t
o
s
eco
nd
proc
essing
sta
ge,
lowe
r
pa
rt
in
F
igure
6.
As
f
unct
ion
a
l
m
od
el
of
the
s
econd
sta
ge
a
m
od
ifie
d
s
pik
i
ng n
eu
r
on
m
odel
from
Sect
io
n
2.3
wa
s u
se
d.
M
od
ific
at
io
ns
co
ns
ist
on
kee
ping
on
l
y
so
m
atic
act
ion
m
od
el
,
rem
ov
in
g
re
fr
act
or
y
per
i
od
subsyst
e
m
,
and
t
re
at
in
g
the
i
nput
sig
nal
as
excit
at
or
y
sy
na
ptic
cu
rr
e
nt
with
pro
per
sc
al
ing
,
res
ulti
ng
i
n
P
ulseGe
n
(
pu
lse
ge
ne
ra
tor)
s
ub
syst
em
bl
ock.
Param
et
ers
we
re
set
to
hav
e
m
ean
out
pu
t
pulse
fr
e
quency
eq
ual
to
co
ntr
ol
value
at
i
nput.
I
nternal
ly
a
nois
e
gen
e
rato
r
was
a
d
de
d,
with
par
am
et
ers
a
dju
sta
ble
th
rou
gh
m
ask
of
the
c
orrespo
nd
i
ng
bl
ock.
Each
pulse
gen
e
rato
r
blo
c
k
will
ha
ve
dif
f
eren
t
see
d
for
noise
ge
ne
rato
r
for
un
c
orrelat
ed
no
ise
betwee
n
sig
nals.
Proc
essing
of
the
first
sta
ge
is
done
by
pa
ssing
in
put
sig
nal
th
rou
gh
c
e
nter
sh
ifte
d
gro
up
of
sel
ect
ed
basis
functi
on,
blo
c
k
1L
_S
BF
f
or
on
e lay
er of
s
qua
re
basis fu
nctio
ns
i
n
Fi
gure
6.
Sp
eci
fic
to
this
cere
bella
r
m
odel
is
that
popula
ti
on
co
de
d
m
os
sy
fiber
s
(
m
f1
in
Fig
ur
e
1)
will
ta
r
get
gr
a
nule
cel
ls,
w
hile sepa
rate set
of
ra
te
co
de
d
m
os
sy fibe
rs
(m
f2
in
Fi
gure
1
)
w
il
l
ta
rg
et
lowe
r
de
ndrite
t
ree of
Go
l
gi
cel
l
[21,
28,
23]
.
Inp
ut
s
ign
al
s
x1
an
d
v1
re
pr
ese
nt
a
ny
sig
nal
t
o
be
pr
ocesse
d
by
c
er
ebell
um
.
Ty
pical
ly
for
se
rvom
oto
r
drive
n jo
i
nt x1
w
il
l be
joint
posit
ion,
wh
il
e
v1 w
il
l b
e
joi
nt sp
ee
d or j
oin
t
acce
le
rati
on
.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N:
20
88
-
8708
Ce
rebell
ar
M
odel
Co
ntro
ll
er
wi
th n
ew
M
ode
l of Gra
nule
...
(
La
v
dim
K
ur
t
aj
)
4299
Figure
5. Re
cepti
ve fie
lds
for t
wo
-
dim
ensional
input s
pace (Al
bu
s
f
ull o
ve
rlai
d
CM
AC
).
Eac
h
input co
ntains
3
la
ye
rs by
3 r
ecepti
ve fie
lds
.
Re
cepti
ve fie
ld
w
idt
h
is
3 q
ua
nta. Fo
r
dim
ension
x
a
l
ay
ers
are
AD
G
, BEH an
d
C
FI,
and lay
ers fo
r dim
ension
x
b
a
re a
dg, beh
and
cfi. In
pu
t at
x
a
dim
ension
will
m
ake acti
ve
receptive
f
ie
lds
D
,
E,
and
F, w
her
eas
in
pu
t at
dim
ension
x
b
will
m
ake acti
ve recepti
ve
fi
el
ds
g,
h, an
d f.
F
or act
ive f
ie
ld
s c
orre
s
pondin
g
t
o
input (
x
a
, x
b
)
th
ere
will
b
e
9
t
wo
-
dim
ensional
receptive
f
ie
lds
, Df,
D
g,
D
h,
E
f,
E
g, E
h,
Ff
,
F
g
dh
e
F
h.
Po
sit
io
ns
of the t
wo
-
dim
ension
al
receptive
f
ie
lds
h
a
ve been
s
hif
te
d
a li
tt
le
f
r
om
their r
eal
pos
it
ion dete
rm
ined
from
corres
pondin
g on
e
-
dim
ension
al
r
ecepti
ve
f
ie
lds
to m
ake th
em
m
or
e d
ist
incti
ve
Figure
6. Mo
de
l for
gen
e
rati
ng
popula
ti
on
a
nd
/
or r
at
e
cod
e
d spi
king
m
os
sy fibers
.
Uppe
r part
generate
s
popula
ti
on
co
de
d
s
pik
i
ng sig
na
l x1_L
1 from
input
sign
al
x1. Bloc
k 1L
_S
BF
ge
ne
rates 6
sig
nals
w
it
h
sq
ua
re
basis
f
unct
ion.
Am
plitu
de
of t
hese s
i
gn
al
s
is
conve
rted
t
o
c
orres
pondin
g
s
tream
o
f
pulse
s
(
s
pik
es
)
with
pu
lse
g
e
ne
rator bl
ock P
ul
seGen L1
. L
ower
p
a
rt
gen
e
rates
rate
cod
e
d spi
king
sign
al
v1_R
from
inp
ut
si
gn
al
v1, by
di
rectl
y feedin
g i
nput sig
nal to
the in
pu
t
of
pu
lse
g
e
ner
at
or b
loc
k Pulse
Ge
n
R
2.5
.
A
veragin
g Cerebel
lar
Model
Con
tr
ol
le
r wi
th
Gra
n
ule Cell
–
G
ol
gi C
el
l
Bui
ldi
ng
Bl
ocks
Final
avera
gi
ng
m
od
el
of
the
cerebell
um
will
hav
e
in
the
f
ir
st
processin
g
la
ye
r
sever
al
gran
ule
cel
l
–
Go
l
gi
cel
l
buil
din
g
blo
c
ks
.
Each
of
them
ge
ne
rates
a
gro
up
of
high
er
-
order
recep
ti
ve
fiel
ds
by
usi
ng
m
ul
ti
plica
ti
on
op
e
rato
r,
for
a
ll
po
ssible
c
om
bin
at
ion
s
fro
m
a
set
of
lo
w
er
-
order
i
nput
sign
al
s.
Bl
ocks
ha
ve
abili
ty
of
m
od
ulati
n
g
gro
up
ou
t
pu
t
act
ivit
y
by
sepa
rate
m
od
ulati
ng
in
pu
t
sig
nal
[
21
]
.
Inpu
t
sig
nal
s
are
functi
onal
eq
uiv
al
ent
of
m
os
sy
fiber
s.
Set
of
m
os
sy
fibb
ers
m
f1
that
ta
rg
et
directl
y
gr
a
nu
l
e
c
el
ls
is
exp
ec
te
d
t
o
be
with
po
pu
la
ti
on
c
ode.
Ot
he
r
set
m
f2
that
ta
rg
et
s
G
olg
i
c
el
l
lowe
r
dend
r
it
e
tree
(m
odulati
ng
i
nput
si
gnal
)
is
exp
ect
e
d
t
o
be
rate
co
de
d.
Sta
nd
a
r
d
CM
AC
is
sp
eci
al
ca
se
of
this
ce
re
bella
r
m
od
el
wit
h
m
od
ulati
ng
i
nput
set
to consta
nt,
ty
pi
cal
ly
1
.
Ou
t
pu
ts
of
bu
i
lding
bl
ocks
w
il
l
be
functi
on
al
equ
i
valent
of
pa
rall
e
l
fibe
r
s.
Ne
xt
sta
ge
i
s
com
po
se
d
from
on
e
or
m
or
e
P
urkin
j
e
c
e
ll
s
blo
c
ks
,
m
od
el
ed
as
sta
ndar
d
per
ce
ptr
on
w
it
h
li
near
act
iv
at
ion
functi
on,
and
a
nu
m
ber
of
syn
aptic
weig
ht
th
at
is
determ
ined
f
r
om
di
m
ensi
on
al
it
y
of
sig
na
l
(p
a
rall
el
fibe
rs)
e
nteri
ng
t
he
blo
c
k.
Pu
r
ki
nj
e
cel
l
bl
ock
s
receive
t
eachin
g
sig
nal
s
(equivale
nt
of
cl
i
m
bin
g
fi
be
rs)
that
ca
n
be
sam
e
or
dif
fer
e
nt
for
al
l
blo
cks
,
or
i
n
any
c
om
bin
at
ion
.
Eac
h
outpu
t
will
hav
e
on
e
ne
uro
n
of
deep
cere
bella
r
nucl
ei
as
la
st
sta
ge
,
m
od
el
ed
as
sim
ple
su
m
m
at
i
on,
recei
v
in
g
ou
t
pu
ts
from
c
orres
pondin
g
Pu
r
ki
nj
e
cel
ls.
All
sign
al
s
m
ay
hav
e
po
sit
ive
a
nd
ne
gative
value
s,
con
t
rar
y
t
o
bio
l
og
ic
al
e
quivale
nt,
but
c
omm
o
n
t
o
m
os
t
arti
fi
ci
al
neural
net
works.
Othe
r
sit
es
of
sy
napt
ic
con
necti
on
t
hat
m
ay
be
ad
justable
[
36]
,
besi
de
par
al
le
l
fiber
-
P
urkin
j
e
cel
l,
are
no
t
m
od
el
ed.
For
bi
olo
gical
ly
consi
ste
nt
structu
ral
m
od
el
fo
r
s
ome
of
them
adjus
ta
bili
ty
would
be
nee
ded
t
o
bri
ng
m
od
el
at
the
op
e
rati
ng
poin
t,
or
s
om
e
int
erd
e
pe
nd
e
ncy
fu
lfil
le
d.
W
it
h
ne
w
funct
io
nalit
y
assigne
d t
o G
ol
gi
cel
l,
i
nclu
di
ng
ne
w f
unct
io
nalit
ie
s
to
ste
ll
a
te
cel
l
and
bas
ket
cel
l
from
[2
1
]
thi
s
cerebell
ar
m
od
el
was refe
rr
e
d
to
as
CM
ACgbs
[28
]
,
[
23]
.
Si
m
ulati
on
m
od
el
of
CM
A
Cgb
s
is
giv
e
n
in
the
Fig
ur
e
7
f
or
si
ngle
ou
t
pu
t.
M
odel
co
ntains
pr
e
processi
ng
blo
c
ks
f
or
ge
ne
rati
n
g
popula
ti
on
co
de
d
m
ossy
fibe
rs
of
sel
ect
ed
s
ha
pe
(s
qu
a
re,
tria
ng
ular,
or
Gau
s
sia
n)
fro
m
inp
ut
sig
nals
, t
ran
s
form
at
io
n com
m
on
f
or
m
os
t
of
these
m
od
el
s. Fu
nctionally
these
blo
ck
s
ar
e
f
_
R
1
f
_
M
1
[
x
1
]
[
v
1
]
P
u
ls
e
G
e
n
R
Uc
P
u
l
s
e
F
r
e
q
P
u
ls
e
G
e
n
L
1
Uc
P
u
l
s
e
F
r
e
q
[
v
1
_
R
]
[
x
1
_
L
1
]
1
L
_
S
B
F
1
x
x
_
S
B
F
Po
p
u
latio
n
Co
d
ed
Inp
u
t Sign
als
Rate Co
d
ed
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
20
88
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
201
8
:
4292
-
4309
4300
thought
t
o
re
side
ou
tsi
de
of
the
ce
reb
el
lum
.
These
blo
c
ks
an
d
gr
a
nule
c
el
l
–
Go
l
gi
cel
l
blo
c
ks
a
re
s
ha
rab
le
com
po
ne
nts
[
23]
.
If
te
ac
hing
sign
al
is
sam
e
for
al
l
P
urkin
je
cel
l
blo
c
ks
t
hey
ca
n
be
joined
t
o
sin
gle
P
urkin
j
e
cel
l
blo
c
k,
m
aking
dee
p
cere
be
ll
ar
nucl
ei
unnecessa
ry.
Stan
dard
CM
AC
fol
lows
t
his
r
ou
te
.
Gainin
g
kn
owle
dge
with
b
et
te
r
ge
ner
al
iz
at
ion
m
ay
re
qu
ire
dif
fer
e
nt
te
achi
ng
si
gn
al
s
f
or
diff
e
re
nt
P
urki
nj
e
cel
ls,
sim
i
la
r
t
o
bio
lo
gical
case
wh
e
re
si
ng
le
cl
i
m
bin
g
fi
ber
(t
eachin
g
sig
nal)
ta
rg
et
s
on
ly
se
ver
al
P
urki
nj
e
cel
ls
and
eac
h
m
at
ur
e
Pu
r
ki
nj
e
cel
l
is
ta
rg
et
e
d
from
sing
le
cl
i
m
bin
g
fibe
r.
CM
AC
gb
s
pr
ese
r
ves
t
his
str
uctu
ral
pro
per
ty
.
G
ranul
e
cel
l
–
Go
l
gi
cel
l
blo
c
ks
ha
ve
f
our
in
pu
ts
m
at
ching
ty
pical
nu
m
ber
of
dendr
it
es
that
gr
a
nule
cel
ls
hav
e.
If
rece
ptive
fiel
ds
of
lo
wer
dim
ension
al
it
y
are
re
qu
ire
d
unuse
d
in
pu
ts
ar
e
ti
ed
to
co
ns
ta
nt
1,
uppe
r
bloc
k
in
Fi
gure
7
.
If
pure
rate
c
od
e
d
sig
na
l
is
t
o
be
pa
ss
ed
t
o
P
urkin
j
e
cel
l
al
l
inputs
can
be
ti
ed
to
c
on
sta
nt,
ch
oice
offe
re
d
by
swi
tc
hes
for t
w
o ne
xt
bl
ocks.
I
n si
m
ulati
on
s
the
re
is
no
nee
d t
o
ha
ve
these
tw
o bl
oc
ks
at
al
l,
but
bi
ologica
ll
y
there
is
no
direct r
oute
fr
om
m
os
sy fibers t
o
P
urki
nj
e
cell
s.
Figure
7
.
Sim
ulati
on
m
od
el
f
or
CM
ACgbs
.
Mod
el
c
onta
in
s thr
ee
GrC
–
G
oC
bu
il
di
ng b
l
ocks
wit
h
m
odulati
ng
input an
d
t
hr
ee
Purk
i
nje cel
ls
with se
par
at
e l
earn
i
ng sig
nals
sit
uated
i
ns
ide
cere
bel
la
r
c
or
t
ex.
O
nly o
ne
dim
ension
is c
onve
rted
t
o popu
la
ti
on c
ode
. Purki
nj
e
cell
s
are s
umm
ed
at
deep cere
bella
r
nuclei
, m
od
el
ed
as
si
m
ple su
m
m
a
t
ion
,
w
it
h i
ts o
ut
pu
t
bein
g
t
otal ce
re
bella
r
c
ontrol
ou
t
pu
t.
PC
: Po
pu
la
ti
on C
od
e
d si
gnal
, R
C:
R
at
e Cod
e
d
si
gnal
2.
6
.
R
obot
Pla
nt
Robot pla
nt w
i
th one
ro
ta
ry li
nk, m
ou
nte
d
on at ar
bitrary
posit
ion
i
ng
a
nd
or
ie
nting
base
, was
us
e
d
as
con
t
ro
l
plant
for
te
sti
ng
ge
ner
a
li
zat
ion
capa
bili
ti
es
of
CM
AC
gb
s
.
Gen
e
ral
dy
nam
ic
s
m
od
el
[37]
de
scri
bing
thi
s
rob
ot in disag
gregate
d
f
orm
[2
1] is
giv
e
n by
foll
ow
i
ng equ
at
ion
[28]
,
)
(
)
(
)
(
)
(
)
(
)
(
)
(
0
0
0
0
0
0
0
1
'
2
1
1
'
2
1
1
'
1
1
1
11
1
1
1
1
1
3
1
1
2
1
1
1
1
1
11
1
z
L
L
y
L
L
x
L
L
L
'
L
s
d
z
r
y
r
x
r
r
g
m
q
G
g
m
q
G
g
m
q
G
q
m
d
q
F
q
F
g
q
G
g
q
G
g
q
G
q
d
(4
)
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N:
20
88
-
8708
Ce
rebell
ar
M
odel
Co
ntro
ll
er
wi
th n
ew
M
ode
l of Gra
nule
...
(
La
v
dim
K
ur
t
aj
)
4301
wh
e
re:
g
x
0
,
g
y
0
,
and
g
z
0
are
pro
je
ct
ion
s
of
grav
it
y
vector
to
0
-
t
h
c
oor
din
at
e
syst
e
m
;
G
1r1
(
q
1
)
,
G
1r1
(
q
1
)
a
nd
G
1
r1
(
q
1
)
represe
nt
gravi
ty
loadings
ca
use
from
ro
bot
li
nk
a
nd
gri
ppe
r;
d
11r
re
pr
ese
nt
s
inerti
al
loadi
ng
s
,
c
on
sta
nt
in
this
case;
F
1d
an
d
F
1s
represe
nt
dy
nam
ic
and
sta
ti
c
fr
ic
ti
on
;
m
L
is
the
lo
ad
m
ass.
Othe
r
sim
i
la
r
te
rm
s
in
(
4)
hav
e
si
m
il
ar
m
eaning
,
bu
t
a
re
relat
ed
to
l
oad
m
ass.
G
a
nd
d
te
r
m
s
are
in
gen
e
ral
nonlinea
r
f
un
ct
io
ns
de
penden
t
on
j
oi
nt
po
sit
io
n.
If
l
oa
d
is
c
onst
ant
a
nd
it
s
ef
fe
ct
s
are
acc
ount
ed
within
li
nk
and
/
or
gri
pper
,
or
th
ere
is
no
load,
four la
st t
erm
s can be
om
itted.
Dynam
ic
s
of
ba
se
m
ov
em
ent
and
r
otati
ons
a
re
no
t
incl
ud
e
d,
ass
um
ed t
o
be
slo
w e
noug
h
to
be t
re
at
e
d
as
sta
ti
c.
Dim
ensio
nalit
y
of
the
in
put
spa
c
e
de
pe
nd
i
ng
on
the
dynam
ical
eff
ect
s
that
will
be
acc
ount
ed
f
or
ranges
from
1
to
7,
bein
g
1
f
or
li
nk
a
nd
gri
pp
er
gra
vity
loadi
ng
only
,
2
f
or
s
ta
nd
a
rd
r
obot
in
fi
xed
base
wi
thout
fr
ic
ti
on
a
nd
3
i
nclu
ding
fr
ic
ti
on
ef
fects,
5
-
6
for
pr
e
vi
ou
s
ca
se
with
diff
e
re
nt
base
r
otati
on
c
onfi
gurati
ons,
a
nd
7
f
or
f
ull
m
od
e
l
includi
ng
var
i
able
m
ass.
Fo
r
pend
ulu
m
-
li
ke
rob
ot
with
r
od
-
li
ke
li
nk
,
gr
i
pper
m
od
el
ed
as
po
i
nt
m
ass, and
incl
ud
i
ng v
isc
ous
f
rict
ion
(4) bec
om
es
,
)
s
i
n
(
2
1
3
1
1
1
1
1
q
F
g
l
m
m
q
l
m
m
d
S
G
S
2
S
G
S
(
5
)
wh
e
re
m
S
an
d
l
s
are
li
nk
m
ass
and
le
ngth
.
Sim
ula
ti
on
m
od
e
l
of
the
rob
ot
[
23
]
was
de
vel
oped
in
Sim
Me
c
han
ic
s
.
As
m
ai
n
co
ntr
oller P
rop
or
ti
onal
-
Der
ivati
ve (PD) c
ontr
oller w
as
us
e
d.
3.
RESU
LT
S
AND A
N
ALYSIS
First
processi
ng
of i
nfor
m
at
ion
e
nteri
ng t
he c
ereb
el
lum
b
y
m
os
sy fibers i
s
do
ne
in
the
granu
la
r
la
ye
r
,
with
ci
rcu
it
s
f
orm
ed
by
gr
an
ul
e
cel
ls
and
G
ol
gi
cel
ls,
and
r
esult
is
passed
to
m
olecular
layer
by
gr
a
nule
cel
ls
axons
for
f
ur
t
he
r
proces
sin
g.
Gr
a
nule
cel
ls
axons
a
re
com
po
se
d
of
two
di
sti
nctive
p
arts:
ascend
i
ng
a
xo
n
an
d
par
al
le
l
fibe
r,
aa
an
d
PF
in
F
igure
1.
Ver
ti
c
al
ascen
ding
a
xon
par
t
sta
rts
from
gr
an
ule
c
el
l
and
asce
nd
s
up
to
so
m
e
le
vel
of
t
he
m
olecular
l
ay
er,
w
her
e
it
bif
ur
cat
es
in
f
orm
of
le
tt
er
T
a
nd
c
reates
pa
ra
ll
el
fiber
s
t
hat
travel
so
m
e
distance
in
both
di
rec
ti
on
s.
O
utputs
f
ro
m
gro
up
of
ab
out
5000
gran
ule
cel
ls
will
m
ake
synaptic
connecti
ons
wi
th
on
e
Go
l
gi
ce
ll
[27,
23]
,
w
hich i
n t
urn
will
inh
i
bit
sam
e
group
of
gran
ule
cel
ls.
This
fee
dbac
k
inh
ibit
io
n
is
do
ne
t
hroug
h
sy
na
ptic
co
nnect
io
ns
i
ns
ide
gl
ome
ru
li
,
Gl
i
n
Fi
gure
1,
where
se
t
m
f1
of
m
os
sy
fibe
rs
m
ake
excit
at
ory
synaptic
connecti
ons
with
gr
a
nule
cel
l
den
dri
te
s.
Sec
ond
route
with
functi
onal
ly
disti
nct
inf
or
m
at
ion
c
om
es
by
s
et
m
f2
of
m
os
sy
fibe
r
s
to
Go
l
gi
cel
l
t
hro
ugh
it
s
lo
we
r
de
ndrite
t
ree
l
dt,
with
fee
dfo
r
ward
inh
ibit
ory
e
ff
e
ct
on
gra
nu
le
cel
ls.
This
w
ould
be
one
i
de
al
iz
ed
el
em
ent
ary
gran
ule
ce
ll
-
Go
l
gi
cel
l
buil
ding
blo
c
k.
3.1.
M
od
ul
at
e
d Higher
Or
d
er Recep
tive
Fie
lds w
ith G
ranu
le
Cell
–
G
olg
i
Cell
Bui
ldi
ng
Bl
oc
ks
Lo
wer
dim
ension
al
inf
orm
ation
at
t
he
i
np
ut
of
the
ce
re
bellum
carried
by
m
os
sy
fi
be
rs
is
first
processe
d
by
gr
a
nule
cel
l
la
ye
r.
It
will
ex
pand
t
he
in
for
m
at
ion
sp
ace
by
seve
ral
or
de
rs
of
m
agn
it
ude
by
gen
e
rati
ng
higher
-
or
der
rece
ptive
fiel
ds
tha
t
ser
ve
as
ass
oc
ia
ti
on
s
for
ne
xt
processi
ng
s
te
ps
,
te
rm
ed
by
Albu
s
as
e
xp
a
ns
i
on
r
ecod
i
ng
[3
]
.
T
ypic
al
ly
in
si
m
ula
ti
on
s
i
nput
s
are
on
e
-
dim
ensio
nal
i
nform
at
ion
a
nd
processe
d
inf
or
m
at
ion
carried
by
a
xons
of
gran
ule
cel
ls
(asce
nd
i
ng
ax
on
s
an
d
paral
le
l
fiber
s
)
at
th
e
ou
t
pu
t
of
gra
nule
cel
l
la
ye
r
are m
ulti
-
dim
ension
al
.
How
e
ff
ect
ive
is
sp
iki
ng
sin
gle
gran
ule
ce
ll
–
Go
l
gi
cel
l
el
e
m
entary
bu
il
ding
blo
c
k
in
gen
e
rati
ng
higher
-
or
der
re
cepti
ve
fiel
d
from
on
e
-
dim
ension
al
i
nputs
is
ex
plored
.
Sim
ulati
on
s
we
re
done
with
m
od
e
l
from
Figure
4.
All
gr
anu
le
cel
ls
in
the
m
od
el
hav
e
sam
e
values
f
or
pa
ram
et
ers
[3
1],
an
d
each
of
them
has
fo
ur
input
s
(b
asal
dend
rite
s).
Accor
ding
to
ideal
iz
ed
m
od
el
,
only
one
Go
l
gi
cel
l
is
us
e
d.
It
receiv
es
excit
at
or
y
s
ynaptic
inputs
from
al
l
gr
a
nu
le
cel
ls
in
the
bl
o
c
k
an
d
sen
ds
in
hib
it
or
y
syna
ptic
outp
uts
to
the
s
a
m
e
gr
an
ule
cel
ls.
All
bu
t
one
relat
ive syna
ptic st
re
ng
t
h gain
bloc
ks
i
n
the
m
od
el
are
set
t
o 1.
The
one
dif
fer
e
nt
is
gain
S
_Gr
C
-
G
oC
set
t
o
1/
(num
ber
of
granu
le
cel
ls
ta
r
ge
ti
ng
G
olg
i
cel
l)
f
or
prop
e
r
scal
in
g
of
sy
na
ptic
stren
gth
[
31
]
.
Inp
ut
sp
ac
e
is
two
dim
en
sion
al
c
om
po
sed
f
ro
m
two
one
-
dim
ension
a
l
inp
ut
sign
al
s.
T
wo
la
ye
rs
of
square
r
ecepti
ve
fiel
ds
are
us
e
d
f
or
ea
ch
i
nput
sig
nal,
sim
il
ar
to
the
on
e
s
how
n
in
F
igur
e
5.
I
n
t
his
way
there w
il
l
be
f
our
gro
up
s o
f
nono
ver
la
pp
i
ng
r
ecepti
ve
fiel
ds
that
carry
in
f
or
m
at
ion
for
in
pu
ts o
f
gr
a
nule
cel
ls.
Lay
ers
a
re
com
po
s
ed
f
r
om
2
r
ecepti
ve
fiel
ds
2
quanta
wide.
Four
c
om
bin
at
i
on
s
,
f
ro
m
9
po
ssible,
wer
e
im
ple
m
e
nted
ta
r
getin
g
4
gran
ule
cel
ls.
Mo
del
from
Figure
6
is
use
d
t
o
ge
ner
at
e
s
piki
ng
m
os
sy
fi
be
rs
from
receptive
fiel
d
act
ivit
y.
To
e
xplo
re
w
ho
le
in
pu
t
s
pace
t
wo
sawto
oth
sig
na
ls
are
us
ed
.
Fre
qu
e
ncy
of
x1
sign
a
l
(hor
iz
on
ta
l
axi
s)
is
se
ver
al
ti
m
es
higher
t
ha
n
t
hat
of
x2
si
gn
al
(
ver
ti
cal
a
xis),
dep
e
ndin
g
on
re
qu
ire
d
ver
ti
cal
reso
l
ution.
Thi
s
sim
ulati
on
use
d
1
Hz
a
nd
0.01
Hz
fr
e
que
ncies.
I
nput
spa
ce
is
sca
nned
in
[
-
8,
8]
rang
e,
but
eff
ect
ive
u
se
d space is i
n [
-
4,
2] r
a
ng
e
, fo
r b
oth
dim
ension
s
.
Figure
8
s
hows
sp
i
king
act
ivit
y
at
outp
uts
of
4
sel
ect
e
d
gran
ule
cel
ls.
D
ots
repres
e
nt
pr
ese
nce
of
s
pi
ke
at
corres
ponding
po
sit
io
n
a
nd
ti
m
e.
Den
sit
y
of
do
ts
is
visu
al
m
easur
e
of
act
ivit
y.
Ce
nters
of
act
ivit
y
for
Evaluation Warning : The document was created with Spire.PDF for Python.