2006
10 h
In e na ional
Wo kshop
on
Cellula
Neu al
Ne wo ks
and
Thei
Applica ions,
Is anbul,
Tu key,
28-30
Augus
2006
3-Laye
CNN
Chip
o
Focal-Plane
Complex
Dynamics
wi h
Adap i e
Image
Cap u e
C.
M.
Dominguez-Ma as,
R.
Ca mona-Galan,
F.
J.
Sainchez-Fe naindez,
A.
Rod iguez-Vazquez
Ins i u o
de
Mic oelec 6nica
de
Se illa-CNM-
CSIC
A .
Reina
Me cedes
s/n,
41012
Se illa,
Spain
e-mail:
[email p o ec ed]
Abs ac
This
pape
p esen s
a
CMOS
implemen a ion
o
a
laye ed
CNN
concu en
wi h
32x32
pho osenso s
wi h
locally
p og ammable
in eg a ion
ime
o
adap i e
image
cap u e.
The
ne wo k
is
a anged
in
wo
laye s
con aining
eedback
and
con ol
empla es,
in e -laye
connec ions
and
p og ammable
a io
o
ime
cons an s.
The e
a e
also
eed o wa d
connec ions
o
a
hi d
laye ,
which
is
as e ,
and
de o ed
exclusi ely
o
combining
he
ou pu s
o
he
o he
wo.
A
mo e
obus
and
linea
mul iplie
block
has
been
employed
o
educe
i egula
analog
wa e
p opaga ion
ough
o
asymme ic
synapses.
Global
and
local
adap a ion
ci cui s
a e
included
on-chip.
The
p edic ed
compu ing
powe
pe
powe
consump ion,
240MOPS/mW,
is
amongs
he
la ges
epo ed,
wha
ende s
his
kind
o
de ices
as
especially
adequa e
o
po able
applica ions
o
a i icial
ision.
Index
Te ms
Vision
chips,
CNN,
pa allel
p ocessing.
I.
INTRODUCTION
Fo
he
mos
o
us
humans,
ision
is
he
dominan
senso y
modali y
in
he
acquisi ion
o
in o ma ion
om
he
en i onmen .
Fo
his
o
be
possible,
na u e
has
de eloped
one
o
he
mos
e icien
de ices
in ended
o
adap i e
image
cap u e
and
eal- ime
image
p ocessing:
he
e ina
[1].
Meanwhile,
he
s uggle
o
b ing
a i icial
ision
o
ai ly
inaccessible
places
con inues.
This
is
mainly
ough
o
he
di icul ies
o
handle
ex ao dina y
amoun
o
da a
con ained
in
he
isual
s imuli
wi h
he
help
o
con en ional
mic op ocesso s.
E en
i
such
da a
low
can
be
managed,
i
is
done
a
he
expense
o
conside able
physical
p o ile
and
ene gy
consump ion.
This
conce n
migh
no
be
a
p oblem
in
machine
ision
applica ions
in
indus ial
en i onmen s.
Howe e ,
in
applica ions
like
obo ic
ision
[2],
senso
ne wo ks
o
ambien
in elligence
[3]
o
e inal
p os hesis
o
he
blind
[4],
powe
e icien
compu a ion
and
he
use
o
he
simples
and
he
leas
ha dwa e
possible
a e
manda o y.
He e
is
whe e
con en ional
digi al
p ocesso s,
wi h
a
se ial
p ocessing
scheme,
ail
o
mee
he
speci ica ions.
As
can
be
seen
in
Fig.
1,
gene al
pu pose
p ocesso s
a e
no
e y
ene gy
e icien .
DSP's
and
ha dwa e
accele a ed
p ocesso s
pe o m
be e ,
bu
he
eal
boos
in
pe o mance
is
ob ained
by
he
This
wo k was
pa ially
suppo ed
by
p ojec s
TIC2003-09817-C02-01
o
he
Spanish
MCyT,
and
N-00014-02-1-0884
o
he
ONR.
F.
J.
Sanchez
is
suppo ed
by
a
g an
o
he
Spanish
MEC.
adap a ion
o
he
a chi ec u e
o
he
na u e
o
he
s imuli.
This
is
qui e
common
in
biological
senso y
o gans,
ha
exploi
he
high
le el
o
pa allelism
p esen
in
agg ega es
o
neu al
cells.
In
o de
o
ealize
an
e icien
VLSI
implemen a ion
o
a ay
p ocessing,
analog
and
mixed-signal
ci cui s
ep esen
a
good
al e na i e.
The
numbe
o
ope a ions
pe
second
in
analog
chips
has
been
calcula ed
assuming
peak
pe o mance
is
du ing
a
con olu ion,
wha
ende s
he
o mula
in
[7]:
(Nadd
+
Np od)Ncells
Z'Con (Nbi s
+
1)
ln
2
(1)
The e o e,
he
numbe
o
OPS
is
he
a io
be ween
he
o al
numbe
o
addi ions
and
p oduc s
ealized
in
pa allel
in
he
chip,
and
he
ime
i
akes
o
he
chip
o
se le
o
he
inal
esul
o
he
con olu ion
wi hin
he
equi ed
accu acy.
Back
o
Fig.
1,
using
analog
ci cui s
a
he
elemen a y
p ocessing
uni s
a oids
A/D
con e sion
a
he
pixel
le el,
and,
o
mode a e
accu acy
equi emen s,
hey
occupy
less
a ea
and
consume
less
powe
han
hei
digi al
coun e pa s.
In
he
ollowing
sec ions
he
CACE2
ision
chip,
he
de ails
o
he
elemen a y
p ocessing
uni ,
and
he
ex ended
ea u es
o
he
chip
a e
explained.
II.
CACE2
SYSTEM
DESCRIPTION
The
CACE2
sys em
a chi ec u e
is
in ended
o
be
implemen ed
in
single
chip,
cons i u ing
a
comple e
ision
Fig.
1.
Compu ing
powe
pe
mW
o
di e en
p ocesso s.
1-4244-0640-41061$20.00
©2006
IEEE
i-/
a
0.01
sys em
on
a
chip
(VSoC).
I s
ope a ion
will
be
con olled
by
an
embedded
mic op ocesso ,
he
CACE2
MCU
(Fig.
2).
A
his
poin ,
his
a chi ec u e
has
been implemen ed
in
2
chips:
image
cap u e
and
ea ly
ision
asks
a e
ealized
by
a
specialized
pe iphe al,
he
CACE2
APAP,
which
is
he
chip
epo ed
he e,
and
an
FPGA
con aining
he
CACE2
MCU.
The
APAP
consis s
in
an
analog/mixed-signal
pa allel
a ay
p ocesso
o
32x32
cells.
Each
cell
is
equipped
wi h
p og ammable
spa io- empo al
dynamics,
local
suppo
o
analog
and
logic
in-pixel
a i hme ic
ope a ions,
local
analog
and
logic
memo ies
and
a
pho osenso
wi h
ex ensions
o
adap i e
image
cap u e
(Fig.
3).
I
ollows
he
a chi ec u e
o
he
CNN-UM
[9]:
an
analog
p og ammable
a ay
con olled
by
signals
common
o
all
o
hem
and
s o ed
in
i s
in e nal
swi ch
con igu a ion
egis e s
(SCR's).
F om
he
poin
o
iew
o
he
ne wo k
opology,
each
cell
inco po a es
wo
nodes
o
a
CNN,
belonging
o
laye s
o
di e en
ime
cons an s,
and
a
hi d
node
o
combining
hei
ou pu s.
This
ne wo k
suppo s
complex
dynamic
phenomena
exp essed
by
a
se
o
coupled
eac ion-di usion
equa ions
[10].
In
his
i s
p o o ype
o
he
sys em,
he
CACE2
MCU
has
been
implemen ed
in
a
FPGA,
oge he
wi h
he
necessa y
pe iphe als
o
boo ing
up
he
sys em,
p og am
and
da a
s o age
and
communica ion.
The
SCR's
o
he
APAP
a e
alloca ed
wi hin
he
add ess
space
o
he
MCU.
The
p og ammabili y
and
con ol
o
he
ne wo k
dynamics,
he
calib a ion
and
biasing
o
he
analog
and
mixed-signal
building
blocks,
he
adap i e
image
cap u e
mechanisms,
a e
con olled
by
he
MCU
ia
he
signals
s o ed
in
he
SCR's.
Access
o
hem
is
di ec ed
by
he
add ess
bus
(A-bus).
The
access
mode,
ei he
eading
o
w i ing,
is
indica ed
by
he
con ol
bus
(C-bus).
The
da a
bus
(D-bus)
is
employed
o
sending
o
ecei ing
pa ame e s.
Each
SCR
mus
be
p ope ly
upda ed
o
manage
he
in a-
and
in e -cell
connec i i y
o
he
p ocessing
elemen s,
he
sequences
o
signals
ha
con ol
he
a ay
ope a ion
and
he
codes
o
he
in e nally
gene a ed
analog
e e ences
employed
o
mixed-signal
ci cui s
ope a ion.
The
a ay
p ocesso
coun s
also
wi h
a
seconda y
da a
bus
(8-bi s
wide)
employed
o
image
I/0.
This
bus
is
di ec ly
connec ed
o
he
sys em
memo y
ia
an
access
con olle
(DMA)
which
is
also
ac i a ed
by
he
MCU
when
equi ed
by
he
so wa e
p og am.
III.
CNN
PROCESSING
UNIT
The
ype
o
signal
p ocessing
ealized
by
he
CACE2
APAP
is
based
on
he
dynamic
e olu ion
o
a
3x32x32
CNN.
This
beha io
is
desc ibed
in
e ms
o
he
inpu
(Uk),
s a e
(Xk)
and
ou pu
(Yk)
a iables.
Each
laye ,
k,
o
he
a ay
ollows
he
e olu ion
law
exp essed
by:
xk()
)]+I[Ak,
Oy
+Bk,
Ou,]+z
d
n
k
(2)
The
symbol
0
s ands
o
he
linea
con olu ion
be ween
he
eedback
and
eed o wa d
empla es,
wi h
he
ou pu
and
inpu
ma ices
o
laye ,
n,
whe e
n
can
be
1,
2
o
3:
[Ak,
(gyn](i,j)
=
E
A
(l,m)y,(i+
1,
j+m)
1=- m=-
[Bk,
(J
u,
](i
j)
E
YB,
(1,
m)u
(i
+
1
j
+
m)
1=- m=-
(3)
whe e
is
he
neighbou hood
adius.
In
his
pa icula
implemen a ion,
c1l
and
'C2
a e
compa able
while
'C3
is
much
smalle
han
he
o he s.
I
he
ull-signal- ange
CNN
model
is
employed
[11],
he
ou pu
and
s a e
a iables
can
be
iden i ied.
In
his
condi ions,
each
ma ix
elemen
in
Eq.
(3)
is
Fig.
2.
Func ional
diag am
o
he
CACE2
sys em
om
he
con ol
uni
o/ om
he
a ay
1/C
~~~~1i
~
~
~ ~
L
Laye -2
F
l;j
Fig.
3.
Concep ual
diag am
o
he
basic
cell.
ob ained
om
he
mul iplica ion
o
he
s a e
(o
inpu )
a iable
by
a
p og ammable
weigh .
These
ope a o s,
esponsible
o
mul iplying
he
s a e
(o
inpu )
a iable
by
a
p og ammable
weigh ,
a e
e med
synapses
o
synap ic
blocks
in
his
con ex .
They
a e
basically
ou
quad an s
mul iplie s
in
which
linea i y
wi h
he
s a e
(o
inpu )
a iable
and
a
symme ic
cha ac e is ic
a e
s ongly
desi ed.
The
e ec
o
a ying
he
p og ammed
weigh s
is
o
modi y
he
ne wo k
dynamics,
and
hus,
changing
he
ype
o
p ocessing
ealized
by
he
a ay.
Con inuing
wi h
he
e olu ion
law,
he e
is
a
losses
e m:
mO
[xkQ(ix)
-1]
+
m
i
Xk
(i,j)>1l
g[xkQ(,
j)]
=lim
m6xk
(i,
j)
i
Xk
(i,Ij)
<
1
(4)
mO
[xk
(ix,j)
+
1]
-m4
i
XkA(i,j)<
and
he
ac i a ion
unc ion,
o
gene a e
he
ou pu :
YkQ,I
)
[xkQ,i )]
=
'i1
(Sj
+mX
i
)m]
(5)
2~
~
~~~klI
m]}C
In
bo h
equa ions,
mL
can
be
0
o
1
o
ha d
o
sigmoidal
ype
nonlinea i y,
espec i ely.
The
physical
ealiza ion
o
he
elemen a y
p ocessing
uni
o
he
CNN
s a s
wi h
he
selec ion
o
he
app op ia e
o ma
o
he
ep esen a ion
o
he
signals.
On
one
side,
ol ages
can
be
easily
deli e ed
o
neighbou ing
a eas
by
connec ing
wi es
o
high-impedance
nodes.
The e o e,
inpu ,
ou pu
and
s a e
a iables
a e
chosen
o
be
ep esn d by
he
ma ices
o
ol ages
Vu,
Vh
and
Ve,
espec i ely.
On
he
o he
side,
signal
addi ion
can
be
easily
ealized
in
he
o m
o
cu en s
wi ed
oge he
o
a
i ual
g ound.
Hence,
he
summands
in
he
second
membe
o
Eq.
(2)
should
be
ep esen ed
by
cu en s.
And
hen,
his
sum
o
cu en s
will
be
in eg a ed
in
he
s a e
capaci o
o
ob ain
he
ins an aneous
alue
o
he
s a e
a iable
ol age:
dV,k
( )
Ck
d
,g[Vk( )]±Z[GA4kn®(DVyn,
±GB,kn(®VU l±I
+1Z
d
n~~~~~~~~~~~~~~~~~~
(6)
As
can
be
seen,
he
elemen s
o
he
eedback
and
eed o wa d
empla es,
Ak4i,j)
and
Bk4ij),
a e
now
p og ammable
linea
ansconduc ances,
GA,k(i,j)
and
GB,kn(i,j),
ha
mul iplied
by
inpu
and
ou pu
ol ages
ende
he
neighbou hood
con ibu ions
in
he
o m
o
cu en s.
Thus,
he
synap ic
block
is
a
ansconduc o
whose
ou pu
cu en
is
p opo ional,
in
he
ideal
case,
o
he
p oduc
o
he
s a e
(o
inpu )
a iable
and
he
weigh .
The
double
ans o ma ion
implici
in
Eq.
(6),
V-I
and
hen
I-V,
allows
o
a
compac
ealiza ion
o
he
p ocessing
node,
achie ing
highe
cell
densi ies,
meaning
an
a ay
size
o
p ac ical
in e es
and,
besides,
a
ole able
ill
ac o .
The
accu acy
o
hese
e ms
is
e y
impo an
o
accomplish
a
co ec
ope a ion
o
he
ne wo k,
since
he
synapse
o se s,
as
well
as
e e y
misma ch
on
ideally
symme ic
weigh s,
a e
in eg a ed
in
he
s a e
capaci o .
P ecisely,
in
he
implemen a ion
o
ou -quad an
mul iplie s,
one
o
he
common
di icul ies
is
o
main ain
he
symme y
wi h
espec
o
he
o igin
o
he
weigh s.
A
misma ch
in
weigh s
ha ing
he
same
absolu e
alue
bu
opposi e
signs
can
modi y
he
dynamic
ou es
o
he
cells
in
he
ne wo k,
ending
in
displaced
equilib ium
poin s,
and
hus,
dis o ing
he
p esc ibed
p ocessing.
The
main
linea i y
conce ns
a e
ound
in
he
V-I
con e sion,
as
linea
cu en
in eg a ion,
and
hus
I-V
ans o ma ion,
can
be
p o ided
by
a ailable
highly
linea
double-poly
capaci o s.
In
his
design,
we
ha e
employed
a
linea ized
OTA
in
o de
o
gene a e
he
uni a y
cu en
con ibu ion.
Though
he
elemen a y
ansconduc o
achie ing
V-I
con e sion
has
a
la ge
numbe
o
ansis o s
han
he
single- ansis o
synapse
in
[12],
ad an ages
in
he
linea i y
wi h
he
s a e
(o
inpu )
a iable
and
symme y
o
he
V-I
cha ac e is ic
jus i y
i s
use.
In
addi ion,
he
suppo ing
ci cui y
can
be
simpli ied
esul ing
in
a
mo e
obus
implemen a ion
inally
wi hou
any
a ea
penal y.
The
schema ics
in
Fig.
4
ep esen
he
co e
o
he
elemen a y
dynamic
p ocesso .
Ope a ing
in
closed
loop
(when
he
swi ch
con olled
by
'Loop'
is
on),
i
implemen s
he
e olu ion
law
desc ibed
by
Eq.
(6).
The
weigh ed
V-I
con e sion
o
he
s a e
ol age
is
ca ied
a
se e al
s ages.
The
single- o-di e en ial
V- o-I
con e sion
is
ealized
by
a
linea ized
ansconduc o
(le -side
o
he
schema ics),
hese
cu en
signals
a e
eplica ed
and
scaled
by
se e al
p og ammable
cu en
mi o s
o
gene a e
he
con ibu ions
owa ds
he
neighbo s
and
i sel
(a
he
cen e
and
igh
sides)
and
he
cu en
signals
om
he
neighbo s
and
sel - eedback
a e
added
and
in eg a ed
in
he
s a e
capaci o ,
when
eedback
loop
is
closed
(by
he
block
a
he
cen e ).
The
ansconduc o
esponsible
o
ans o ming
he
s a e
capaci o
ol age
V,
in o
a
di e en ial
cu en
is
a
sou ce
degene a ed
di e en ial
pai
wi h
diode-connec ed
loads.
I
is
based
on
a
linea ized
OTA
[13].
The
ope a ion
o
his
ci cui
Fig.
4.
Schema ic o
he
linea ized
OTA
and
synap ic
blocks
alone
is
inhe en ly
symme ic
i
wo king
in
ully-di e en ial
mode,
ep esen ing
an
enhancemen
om
wha
ha e
been
achie ed
by
p e ious
implemen a ions.
This
symme y
hough
is
b oken
by
using
a
single-ended
inpu
ol age,
bu
s ill
he
esul ing
V-I
cha ac e is ic
main ains
symme y
le els
beyond
hose
o
o he
implemen a ions.
The
bene i s
o
a
di e en ial
ep esen a ion
we e
no
signi ican
o
be
wo h
handling
wi h
double
capaci o
a ea
and
a
complex
signal
ou ing.
The
implemen a ion
o
he
weigh s
is
based
on
geome ical
ela ions
be ween
ansis o s.
This
has
he
ad an age
o
being
less
in luenced
by
p ocess
pa ame e
a ia ions
bo h
in e -
and
in a-die.
I
has
also
he
d awback
o
only
pe mi ing
he
use
o
a
disc e e
se
o
weigh
alues,
namely
-4, -2, -1,
0,
1,
2
and
4.
Opposi e-sign
con ibu ions
a e
ob ained
by
c ossing
he
wi es
con eying
he
cu en s
o
he
collec ing
nodes,
hus,
achie ing
by
a chi ec u e
a
symme ic
ope a ion.
Finally,
he
sum
o
all
he
cu en s
coming
om
he
neighbo hood
is
injec ed
in o
he
a ge
s a e
capaci o .
Bu
4lo16A
4
,YW=-4.,
2
W=2
IWI0
2
W
'.
WI;,.,-
-2
W=2
W44
-0.5
-0.4
-0.3
-0.2
-0.1
0
0.1
0.2
0.3
0.4
0.5
V
S a e
a iable
ol age
(Vin-Ve )
Fig.
5.
Ou pu
cu en
s.
s a e
ol age
o
he
OTA-based
synapse.
be o e
ha ,
di e en ial
o
single-ended
cu en
con e sion
is
ealized
wi h
he
help
o
a
cu en
mi o .
I
is
impo an
o
men ion
ha
he
achie able
ou pu
esis ances
using
sel -
biased
o
ex e nally
biased
Cascode
cu en
mi o s
a e
no
su icien
o
ensu e
he
necessa y
independence
om
he
ou pu
ol age.
In
o he
wo ds,
he
esul ing
e o
in
he
copied
cu en ,
because
o
he
ini e
ou pu
esis ance
o
he
mi o ,
was
beyond
he
p edic ed
e o s
due
o
pa ame e s
misma ch.
The e o e,
gain
boos ing
o
he
Cascode
de ices
is
needed
o
educe
his
e ec .
The
accu acy
o
he
cu en
eplica ion
in
his
mi o
is
c ucial
o
achie ing
he
equi ed
linea i y
and
symme y
in
he
V-I
cha ac e is ic.
Also,
we
ha e
employed
0.5/2.0
ansis o s
o
he
cu en
mi o s,
ensu ing
enough
ma ching.
As
a
esul ,
he
ou pu
cu en
has
a
high
linea i y.
The
ansconduc ance
ela i e
e o
in
la ge
signal
is
kep
below
0.7%.
Conce ning
he
symme y
o
he
cha ac e is ic,
he
di e ence
o
he
ou pu
cu en s
co esponding
o
weigh s
wi h
he
same
absolu e
alue
bu
opposi e
sign
is
ze o
on
a e age
because
o se
cancella ion,
he
s anda d
de ia ion,
ob ained
by
Mon e
Ca lo
simula ion,
being
2%
o
he
absolu e
alue
o
he
indi idual
cu en s.
Compa ed
o
p e ious implemen a ions,
in
e ms
o
linea i y
o
he
V-I
cha ac e is ic
o
he
mul iplie s,
his
ci cui
pe o ms
one
o de
o
magni ude
be e
o
compa able
a ea
and
powe
consump ion.
This
is
no
always
equi ed
o
he
co ec
ope a ion,
i.
e.
con e gence
o
he
ne wo k
dynamics
o
he
co ec
equilib ium
poin s,
bu
decisi e
o
linea
di usion.
Func ional
ope a ion
o
he
comple e
cell,
including
he
local
memo ies
and
he
swi ching
ee
ha
allows
communica ion
wi h
he
ou side
o
he
a ay,
has
been
e i ied
by
simula ion.
A
small
ne wo k
composed
o
3x3
cells
ac ually
a
5x5
ne wo k
i
he
bounda y
cells
a e
conside ed
has
been
p og ammed
o
implemen
di e en
image
p ocessing
empla es.
Fig.
6(a)
displays
he
s a e
a iables
o
he
cells
o
he
3x3
ne wo k
when
p og ammed
o
ealize
connec ed
componen
de ec ion
in
he
ho izon al
di ec ion.
F om
he
simula ions,
i
can
be
seen
ha
he
ime
cons an
o
he
cells
is
designed
o
be
unde
I
OOns.
IV.
EXTENDED
CHIP
FEATURES
A.
T ue
esis i e
g id
One
o
he
mos
use ul
ools o
ocal-plane
image
p ocessing
is
he
di usi e
p opaga ion
o
he
pixel
alues.
This
is
achie ed
by
a
Gaussian
lowpass
il e ,
which
o
a
disc e ized
image
g id,
can
be
ealized
by
he
con olu ion
o
he
o iginal
image
wi h
a
spa ial
mask.
The
e olu ion
law
implemen ed
a
e e y
node
o
he
CNN
can
suppo
his
di usi e
dynamics
by
app op ia ely
se ing
he
co ec
in e connec ion
weigh s
in
he
eedback
empla e.
The
main
d awback
o
implemen ing
his
ope a o
and
o he s
using
symme ic
weigh s,
in
a
VLSI
s uc u e
designed
o
ully-
p og ammable
CNN
dynamics
is
misma ch
in
gene a ing
cu en
con ibu ions.
Because
o
he
local
compu a ion
o
he
con ibu ions
o
he
neighbo hood,
he
amoun
o
cu en
being
injec ed
om
cell
C(i,j)
in o
neighbo ing
cell
C(i+lj),
o
ins ance,
does
no
ma ch
in
absolu e
alue
he
cu en
being
injec ed
om
cell
C(i+lj)
back
in o
he
s a e
capaci o
o
cell
C(ij).
The
consequence
o
his
is
easy
o
de i e,
he
supposedly
symme ic
di usion
is
con e ed
in o
an
un uled
p opaga ion
o
he
pixel
alues.
Special
ca e
has
been
pu
in
coun e ac ing
he
e ec s
o
misma ch
by
design.
This
is,
esizing
he
ansis o s
in
o de
o
a oid
excessi e
de ia ion
om
he
nominal
in
he
gene a ion
o
he
uni a y
cu en
con ibu ions.
Apa
om
his,
he
p o o ype
chip
includes
a
ue
esis i e
g id
concu en
wi h
he
CNN
a ay.
Each
cell
con ains
wo
esis o s,
made
o
high- esis i i y
poly-Si,
ha
can
be
lhime
(§)
(a)
connec ed
o
he
s a e
capaci o
in
o de
o
o m
a
ec angula
esis i e
g id.
The
ime
cons an
o
his
g id
is
be ween
0.2-
1.0 s,
and
i
is
no
co ela ed
o
he
CNN
ime
cons an ,
nei he
can
be
con olled
by
he
use .
The
ope a ion
o
he
esis i e
ne wo k
is
illus a ed
in
Fig.
6(b).
He e
he
3x3
a ay
o
cells
is
p og ammed
o
e ol e
wi h
null
empla es,
using
only
he
g id
o
poly
esis o s.
In
less
han
2j s,
he
s a e
ol ages
o
all
cells
con e ge
o
he
global
a e age.
This
is
no
as ,
and
i
becomes
wo se
when
he
size
o
he
ne wo k
inc eases,
bu
i
is
con enien
om
he
poin
o
iew
o
he
con ol
o
he
algo i hm
o
coun
wi h
a
di usion
mechanism
ha
uns
slowe
han
he
swi ch
con igu a ion
upda ing
signals.
B.
Adap i e
image
cap u e
Adap i e
image
cap u e
in
he
CACE2
APAP
is
based
in
he
local
and
global
con ol
o
he
pho osenso s'
gain.
Ope a ing
in
pho ocu en
in eg a ion
mode,
he
ol age
ep esen ing
he
alue
o
he
pixel
depends
on
he
in eg a ion
ime,
i.
e.
o
he
same powe
o
he
inciden
ligh
o e
he
senso
su ace,
a
la ge
in eg a ion
ime
will
allow
he
same
pho ogene a ed
cu en
o
discha ge
he
sensing
capaci ance
o
a
longe
ime,
esul ing
in
a
la ge
ol age
excu sion
o m
he
ese
alue.
In
his
chip,
each
pixel
has
a
ese
ansis o
go e ned
ei he
by
a
global
signal
au oma ic
adap a ion
o
he
in eg a ion
ime
is
o
o
by
a
compa a o
d i en
by
a
local
ese
con ol
ol age
and
a
global
ime-e ol ing
e e ence.
The
local
suppo
o
his
compa ison
is
explained
in
[14].
I s
main
unc ion
is
o
adap
he
local
gain
o
he
pho osenso .
As
hey
a e
in eg a ing
senso s,
his
gain
adjus men
is
achie ed
ia
he
adap a ion
o
he
local
in eg a ion
ime
acco ding
o
a
locally
de i ed
ol age
le el.
In
o de
o
do
ha ,
he
global
e e ence
will
be
an
in e se
ol age
amp
ha
is
deli e ed
o
e e y
senso
in
he
a ay.
When
he
in e se
amp
c osses
wi h
nega i e
slope
he
lime
(s)
(b)
Fig.
6.
Simula ion
o
he
e olu ion
o
a
educed
(3x3)
a ay
wi h
local
memo y
upda e,
o se
cancella ion
and
ese
p ocesses.

local
h eshold,
he
in eg a ion
o
he
pho ocu en
s a s.
In
his
way,
he
da ke
he
pixel
hen
he
la ge
in eg a ion
ime
ha
will
be
alloca ed
o
ha
pixel
o
he
nex
cap u e
he
algo i hm
elies
in
he
co ela ion
be ween
he
alues
o
he
same
pixel
in
di e en
ames
in
a
sequence.
Co espondingly,
he
b igh e
he
pixel,
he
less
ime
i
will
ha e
in
he
nex
cap u e.
Conce ning
he
global
adap a ion
mechanism,
he
in e se
amp
is
cen e ed
in
a
p edic ed
a e age
in eg a ion
ime
alue.
The e o e,
i
he
p e ious
image
cap u e
esul ed
in
an
o e -exposed
pic u e,
he
a e age
ol age
will
be
below
he
middle
poin
o
he
pixels'
ol age
ange.
I
he
p e ious
image
is
unde -exposed,
he
a e age
ol age
will
be
abo e
his
poin .
The
algo i hm
p og ammed
in o
he
chip
co ec s
he
ime
ex en
o
he
amp
acco dingly
in
o de
o
ha e
smalle
exposu es
o
b igh ness
sa u a ed
images
and
la ge
exposu es
o
ex emely
da k
pic u es.
This
is
achie ed
by
compa ing
he
a e age
ol age
o
he
pixels
wi h
uppe
and
lowe
h esholds.
I
he
esul ing
a e age
alls
be ween
hese
h esholds,
he
only
co ec ions
in oduced
a e
due
o
local
adap a ion.
I
he
ol age
alls
abo e/below
he
uppe /lowe
limi ,
a
digi al
ci cui
igge ed
by
hese
compa a o s,
co ec s
he
equency
di ision
ealized
on o
he
sys ems
mas e
clock,
employed
o
gene a e
he
in e se
amp,
in
he
p ope
sense.
This
ends
in
a
wide /na owe
amp
shape
un il
he
a e age
pixel
ol age
alls
be ween
he
wo
h esholds.
V.
CHIP
DATA
AND
CONCLUSIONS
The
p o o ype
chip
has
been
designed
and
ab ica ed
in
a
CMOS
0.35p m.
The
die
size
is
7.6mm
x
7.6mm.
Fig.
7
displays
a
mic opho og aph
o
he
chip.
Table
I
shows
a
su ey
o
chip
da a.
These
ea u es
a e
p edic ed
om
he
simula ion
esul s.
The
chip
in
now
unde
es ,
in
o de
o
con i m
he
expec ed
pe o mance.
REFERENCES
[1]
D. H.
Hubel,
Eye,
B ain
and
Vision.
Scien i ic
Ame ican
Lib a y,
No.
22.
W.
H.
F eeman
and
Co.,
New
Yo k,
1995.
[2]
T.
Makimo o,
T.
T.
Doi,
"Chip
Technologies
o
En e ainmen
Robo s
-
P esen
and
Fu u e".
In .
Elec on
De ices
Mee ing,
pp.
9-16,
Dec.
2002.
[3]
E.
Aa s,
R.
Roo e s,
"IC
Design
Challenges
o
Ambien
In elligence",
Design,
Au oma ion
and
Tes
in
Eu ope
(DATE),
pp.
2-7,
Ma ch
2003.
[4]
Eyal
Ma gali
e
al.
"Re inal
P os hesis
o
he
Blind",
Su ey
o
Oph halmology,
Vol.
47,
No.
4,
pp.
335-
356,
July-Augus
2002.
[5]
T.
Komu o,
I.
Ishii,
M.
Ishikawa,
A.
Yoshida,
"A
Digi al
Vision
Chip
Specialized
o
High-Speed
Ta ge
T acking".
IEEE
T ans.
on
Elec on
De ices,
Vol.
50,
Np.
1,
pp.
191-199,
Jan.
2003.
[6]
P.
Dudek
and
P.
J.
Hicks,
"A
Gene al-Pu pose
P ocesso -pe -Pixel
Analog
SIMD
Vision
Chip".
IEEE
T ansac ions
on
Ci cui s
and
Sys ems-I,
Vol.
52,
No.
1,
pp.
13-20,
Jan.
2005
[7]
G.
Li ian
e
al.,
"A
1000
Fps
A
128*128
Vision
P ocesso
Wi h
8-Bi
Digi ized
I/0".
IEEE
Jou nal
o
Solid-S a e
Ci cui s,
Vol.
39,
No.
7,
pp.
1044-1055,
Jul.
2004.
TABLE
I
CMOS
p ocess
0.35p m
No.
o
CNN
cells
3x32x32
Die
a ea
7.6mm
x
7.6mm
A ay
a ea
6.2mm
x
6.2mm
P ocessing
cell
size
176p m
x
176p m
Cu en
consump ion
pe
cell
300 Ag3.3V
Weigh
esolu ion
7-8b
Image
esolu ion
8b
I/0
a es
10MHz
CNN
ime
cons an
below
1
OOns
[8]
R.
Ca mona
e
al.,
"A
Bio-Inspi ed
2-Laye
Mixed-Signal
Flexible
P og ammable
Chip
o
Ea ly
Vision",
IEEE
T ans.
on
Neu al
Ne wo ks,
Vol.
14,
No.
5,
pp.
1313-1336,
Sep.
2003.
[9]
T.
Roska
and
L.
0.
Chua:
"The
CNN
Uni e sal
Machine:
An
Analogic
A ay
Compu e ".
IEEE
T ans.
on
Ci cui s
and
Sys ems-II,
Vol.
40,
No.
3,
pp.
163-173,
Ma ch
1993.
[10]
Cs.
Rekeczky,
T.
Se ano,
T.
Roska
and
A.
Rod iguez,
"A
S o ed
P og am
2nd
O de /3-Laye
Complex
Cell
CNN-UM".
In .
W.
on
CNN's
and
Apps,
pp.
219-224,
Ca ania,
I aly,
May
2000.
[11]
S.
Espejo
e
al.,
"A
VLSI
O ien ed
Con inuous-Time
CNN
Model".
In .
J.
o
Ci cui
Theo y
and
Apps,
Wiley.
Vol.
24,
No.
3,
pp.
341-356,
May-June
1996.
[12]
R.
Dominguez-Cas o,
S.
Espejo,
A.
Rod iguez-Vazquez
and
R.
Ca mona,
"Fou -Quad an
One-T ansis o -Synapse
o
High-Densi y
CNN
Implemen a ions".
In .
Wo kshop
on
CNNs
and
hei
Apps.
(CNNA'98),
pp.
243-248,
London,
UK,
Ap il
1998.
[13]
F.
K ummenache
and
N.
Joehl,
"A
4-MHz
CMOS
Con inuous-Time
Fil e
wi h
On-Chip
Au oma ic
Tuning".
IEEE
J.
o
Solid-S a e
Ci cui s,
Vol.
23,
pp.
750-758,
June
1988.
[14]
R.
Ca mona,
C.
M.
Dominguez-Ma as,
J.
Cuad i,
F.
Jimenez-Ga ido
and
A.
Rod iguez-Vazquez,
"A
CNN-D i en
Locally
Adap i e
CMOS
Image
Senso ".
In .
Symp.
Ci cui s
and
Sys ems
(ISCAS'04),
Vol.
V,
pp.
457-460,
Vancou e ,
Canada,
May
2004
Fig.
7.
Mic opho og aph
o
he
CACE2
p o o ype.