Biomimetic Pattern Recognition Theory and Its Applications

puppypompΤεχνίτη Νοημοσύνη και Ρομποτική

14 Νοε 2013 (πριν από 3 χρόνια και 8 μήνες)

47 εμφανίσεις

Chi nese J ournal of E l ectr oni cs
Vo113 No3J uly 2004
B i omi met i c P at t er n R ecogn i t i on T heor y
and I t s A p p l i cat i on s﫵
WANG Shoujue and ZHAO Xingtao
(Laboratory of Art cial Neural Networks I nstitute of Semiconductors
Chinese Academy of Sciences Beij ing 100083 China)
A bst ract  B iomi met i c pat t ern recognt i on (B P R ) 
w hi c h is b ased o n co gn it i on"i nst ead of c la ssi fi c at i on 
i s m uc h c loser t o t he f un ct ion of h u man b ei ng  T h e ba-
sis of BP R is the P r inciple of homology- cont inuity (P Hc )
w hi ch mean s t h e d i f f er enc e bet w een t w o sam p les of t he
sam e cl ass m ust b e g r ad ua ll y ch an ged  T he a im of B P R i s
t o fi n d an opt i mal cov er i n g i n t h e f eat ur e sp ace w h ic h em
p has iz es t he s im i l ar i ty  am on g h omo log ous gr ou p mere -
b er s  r at her t h an d i v isi on  in t r adi t i on al pa t t er n r ec og-
n i t io n  Some ap p l i cat i on s of B P R ar e su r v ey ed  i n w h i ch
t h e r es ul t s of B P R ar e mu ch b et t er t h an t he r esu l ts of
S u p po r t V lect or M a ch in e A n ov el ne ur o n mo de1 H y p er
sausage neuron (H SN ) i s shown as a ki nd of cover i ng uni t s
i n B P R T he mat h emat i ca l d escr ip t i on o f H SN is gi v en an d
t h e 2d im ensi on al d i sc r i m in ant b o un d ar y of H S N is sh ow n 
I n t w o sp ec ial c as es i n w hi c h sam p l es ar e d i st r i b ut ed i n
a l i ne segm ent an d a c ir cl e b ot h t h e H S N net w or k s an d
R B F net w o r k s ar e u sed f or c ov er i ng  T h e r esu l t s sh ow t h at
H S N n etw or k s act b et t er t ha n R B F net w or k s i n gen er al i za-
t i on  esp ec ial l y fo r smal l sam pl e s et  w h ich ar e co nson an t
w it h t h e r esu lt s of t he ap p l i cat i on s of B P R  A n d a b r i ef
ex p l anat io n of t h e H SN net w or k s  ad va nt ages i n c ov er i ng
gener al d ist r i bu t ed samp l es i s a lso gi v en 
K ey w or d s  P at t er n r ec ogn i t i on B io mi m et ic  N eu r al
net w or k  H y p er sau sage n eu r on 
I  I nt ro d u ct io n
P att er n r ecogni t i on h as b een develop ed for dozen s o f year s
and many th eor i es have mu shr oomed  A ll of t h ese t h eor i es
ar e bas ed on a st at ist i cal mod el in wh i ch a d eci sion r u le was
defi ned wi t h t he ai m to cl as sify t wo k inds o f samples  I n t hese
t heor i es pat ter n r ecogn it i on i s eq uival ent t o di vi sion  of d i f-
fer ent p att er n s
Biomimetic pattern recognition(BPR) was fi rst introduced
by Academician Wang Shouj ue in 2002 I n this theorypat
t er n r ecogn it i on i s bas ed on cogn it i on i nst ead o f d iv isi on 
I n anot her wor d B P R emph as iz es on cogn it i on of al l samp le
clas s es one by one r at her th an cl as si fi cat ion of many ki nds
of sampl esl 1

B ecause th i s t h eor y i s much cl oser t o t he cl as 
sifi cat ion fun ct i on of human b ei ng it is call ed B i omi met i c
P att er n R ecogn it i on  B y now  t h is novel t heory h as b een
adopt ed i n ma ny app li cat i ons and obt ai ned gr eat successes 
I I  B i om i m et i c P at t er n R ec ogn i t i on
W hen a hum an bein g p er for ms r ecogni t ion he p ut s par t i c
ul ar emph as is on cogn it i on a nd onl y consi der s d i st inct ion 
ser i ously i n ver y few- cases t
 H owever  t h e tr ad i t ional P at 
t er n R ecogn it i on on ly pai d at tent io n t o d ist i nct i on an d over 
l ooked th e concept of cogn it i on T he new n ovel met hod t hat
con cent r at es pat t er n r ecogni t ion o n cogni t i on  B iomi met ic
pattern recognition(BPR)was proposed in Ref_[1]The word
B iom imet ic emp has iz es t hat t h e st ar t i ng p oint o f t he fun c
t ion and mat hemat ical model of pat t er n r ecogn it i on is t he con 
cept of C ognit ion wh ich i s much cl oser to th e cl as si fi cat i on
fun ct i on of hum an b ei ng  B ut i n r esp ect of achi ev in g ap pr oach 
t he new met h od d oesn t concr et e t he concep t of B i om imet ic
I n r eal wor l d  if t wo sampl es of t h e same cl as s ar e not ex 
act ly t he samet he difi er ence between t hem must be gr adual l v
changed So a grad ual ly ch angi ng sequ ence exi st s b et ween t h e
two samp les and ever y samp le i n t he sequence bel on gs t o t he
sam e clas s T h is pr i ncip le of cont inu it y amon g homol ogous
samp l es in featu r e space is cal l ed t he P r i nci pl e of homol ogy
continuity (P HC 1ve can get the mathematical description
of P H C  I n feat ur e space we suppose t hat set A is a poi nt
set i nclu di ng al l samp les in class A  I f z Y  A an d  > 0 ar e
given t her e must ex ist set B
(1)
A l t hough t h e pr i ncip le of ho mo logyco nt i nu it y is n ot avai l abl e
i n t h e t ra dit iona l p at t er n r ecogn it i on and l ear ni ng t heor yi t is
a t r ut h in t h e r eal wor l d [
 A nd i t i s a k in d of pr i or knowl
edge of sampl es di str ib ut ion in t he bi omi met i c p at t er n r ecog
ni t i on  T r ad it i onal pat ter n r ecogni t ion ai med t o get t he opt i-
mal cl as si fi cat i on of di ff er ent ki nds of samples i n t he featu r e
sp ace H owever  t h e b iomi met ic p att er n r ecogn it i on i nt en ds
to fi nd t he opt i mal conn ect i on of h om ologous s amp l es in t h e
feat ur e sp ace T he b as ic poi nt o f B P R is t o anal yze t he r el a-
t ion sh i ps bet ween t r ai ni ng sampl e p oint s i n feat ur e space an d
t he P H C makes t h is bas i c poi nt avai la ble
T hr ough t h e pr i nciple of homol ogycont inu it y we know
t hat al l h om ologous samp les o f cl as s A mu st d ist r i but e in a
M a nuscri pt R ece ived Sept  2003 A ccept ed F eb 2004
f
<
+
P
A
= C
M
 
m


n
z
= 
=

ﺬ http://www.cqvip.com

374 Chinese J ournal o1 Electronics 2004
connect ed r egion i n t he feat ur e space T h is r egi on is wr it t en
aS P  C onsideri ng t he smal l di st urbancesal l t he sampl es near
P ou ght t o b e consi der ed as sampl es of class A So t h e ai m of
B P R is t o fi nd an appr opr i ate cover i ng i n t he feat ur e space of
cl as s A  W d efi ne
d(z P ) = m

i
P
nd(z )
whi ch i s t h e di st an ce bet ween t h e v ect or z a nd set
t he per fect cover i ng set of class A i s

{xld(xA) )
(2)
P  T hen
(3)
wher e 溺 is t h e d ist ance t hr esh old  I n nd imensi on al feat ur e
spacet he cover i n g set P a has a ndimensi on al complex shape
wh ich separ at es t he whol e space i nt o 2 par t s on e is class A 
an d t he ot h er is n ot  T h er efor e pat t er n recogn it i on wit h B P R
is to j udge whether the mapping in feature space of an obj ect
b el ongs t o t he cover i ng set P a or not 
I I I  H yp er Sausage N eural N et work
Because i t is i mpossi bl e t o col lect al l t he samp l es in clas s
A  t he p er fe ct cover in g set P a can not b e const r uct ed  Bu t
we can fi nd a un i on of many si mple un it s t hat is appr oxi mat e
t o t he per fe ct cover in g set  T h er efor e pat t er n r ecogni t i on
with BPR is to j udge whether a point belongs to at least one
un it or n ot 
A n eur on uni t can form a compl ex closed shape[2
 For ex 
ampl ea R B F neur on const ru ct s a hyp er sp her ean d a D oub le
weights neuron (DWN) can construct both the hyperellipse
and some ot her mor e comp l ex shap es[2]

So t h e ar t ifi ci al n eu 
r al net wor k i s a ap pr opr i at e choi ce t o const r uct t he cover i ng
set in B P R [11

A new kind of neuronHyper sausage neuron (HSN)was
proposed in Ref_[4]which constructs a sausage like shape in
fe at ur e space T h e di st ance bet ween z and th e l in e segment
i is
f
d2(z ﷟)={  lIxz l

Xl }
  z l
q(xz1x2) =(z  X1) fz I  x2)
lXl  x2l
T hen a H yp er S ausage N eur on can b e wr it t en as
(5)
 d (z i ﻷ ) 
Hsw(z)=sgn(2_  _05) (6)
is al so sh ow n i n F ig1 I t can be seen th at t he H SN s cover i ng
ar ea i s lar ger t han D R B F s wh en r 0  1 t he cover in g ar eas
of H S N and DR BF ar e pr ox i mate when r 0  1 B oth t he COV
er i ng r egion of H SN and of D R B F ar e app r oxi mat e t o cir cl es
when rn  1but the equivalent radius of DRBF is 4 2 times
H S N s  So w it h t he same r ad iu s par amet er t h e H SN can act
bet ter t han R B F neu r on s for cover i ng t h e d i st r i bu t ion r egi on
of a cert ai n cl ass i n fe at ur e space
I V  M at hem at i c M od el a nd A n al y si s
Biomimetics pattern recognition (BPR) emphas izes Cog
ni t i on r at h er t hat Di vi sion  B P R is to fi nd t he opt imal
cover i ng r at her t h an t he op t im al sep arat i ng hyp er su r face I n
B P R onl y t he dat a of one cl ass ar e avai l abl e i n l ear ni ng pe
r i od  T h er efo r e th e m at h emat i c mod el i s much d ifi er ent fr om
t he mod el of t r adi t ion al p at ter n r ecogn i t ion met ho ds  I n th is
pap er onl y th e dis tr i bu t ion t hat is top ologi call y h omomorp h i
cal t o a l ine is di scussed
I deallywe suppose that (1) ignoring the dist urbanceall
t h e samp les of clas s A ar e wel ldis tr i bu t ed p oi nt s i n a l in e P 
w r it t en as
p )    (8)
0 P
where lIP ll is the length of the line P(2) the probability of
th e d ist ur ba nce fo r a vect or fo ll ows t he nor mal d ist r ib ut i on 
w r it t e n as
p (Ul ￳ ( )
Eq(3) given a sample set Aan approximate covering set Pa
can be constructed as Eq(31By the suppositions listed above
we can get th e p rob abi l it y d ist r ib ut i on densi ty funct i on 
p( )= p (ylx)p (x)dx (1o)
  P
S o t h e cor r ect r at e and t he r e ect ion r at e can be cal cu lat ed 
H yper Sau sage n eur ons were us ed t o const r uct t he cover i ng set
of t wo sp eci al cas es a l in e s egment and a cir cle R BF neur ons
wer e also us ed in t hese two cas es as cont r oI exp er i ment 
1 L i n e se gm en t
W it hou t l os in g genera li zat i on  we cons ider a sampl e s et
with only two samplesA(f00) and B(100) in the
ndi mensi on al fe at ur e space So
T h i s f unct ion i s a di scr i mi nant funct ion fo r pat t er n recog and
ni t ion  I f t he fu nct i on r esul t is negat i ve t he vect or b e
lon gs t o t his cl as s  ot her wi se
Xl  ( 100) and x2 =
i t d oesn t  We sup pose th at
(100 ) The discriminant
bou ndar y of t he H y per sausage neu ron i n 2- d imensi on al case
is shown in F i g1 A s t he contr as t  th e d iscr i mi nant bo und ar y
of the DoubleRBF (DRBF) neurons
 ( )- sg (2TIX--Xll2fD +2 _05) (7) RBF( )sgnf2 +2 o5) (7
p ) 
1

  (11)
0 AB
p( )=p ( I )p ( )d
=
 ( ( )r( ))
We constr u ct a Hyper Sau sage neu r on t o conn ect t h e two
points as Eq(6) and DRBF neurons as Eq(7)
2
 ﷒

O 
< >
) ) o 2 2
i
g g

2 2 2
ﺬ http://www.cqvip.com

Bi omi met i c Pat t ern Recognit i on T heory and I ts A ppl icati ons 375

1
L 5
 

2

n
n
54'v- - ~  0
3 2  0

5-   
F ig 1T he discr iminant boundary of tt SN and
di mensional case f or diff erent r 0
So t he cor rect r ecogn it i on rat e of H SN is 
窱hs )=  p(ﯚ
f HSN( ) O
an d t he cor rect r at e of DR B F i s
PDRBF(r0) p(x)dx
D R B F i n 2
T he k sampli ng poi nt s in t he ci rcle wit h t he same i nt erval are
denoted aS{s z) 1An extra point 8k+l 81 is appended in
the sample set There are k line segments (HSNs) that connect
t he  + 1 poi nts  T hen t he di scr i mi nant fun ct i on i s
HSN( (sgn(2Td 5)) (17)  (  ﷲ) 
A R B F net work is const ru ct ed by t he sample set and t he di s
cri mi n ant fu nct i on is 
RB ( )=sgn(k 2 8i 2_05)  I  I 
(13) For a certain lro the ideal correct rate is
P(ro) =  p(x)dx =99
d d(X G(f )) TO
For difi erent fro is selected to keep SN = 99 and
PDRBF(r o) is calculated The results are shown in F ig2and
i t can be seen th at t h e cor r ect rat e of D R B F appr ox imat es t o
the correct rate of HSN when f r o  l and t he correct rate
of DRBF falls down quickly when lro increasesSo the gen
eral izat ion abi li t y of HSN i s st ronger t han DR B F when t he
dist ance bet ween sampl es i s lar ge
0
0
0
1ro
F ig2T he correct rate of HSN and DRBF for diti erent 2ro
For diff erent Iro is selected to keep 窱HSN  99 
T he top l ine indi cates t he corr ect rates of H SN whi le
the ot her curv es fr om up t o dow n  are t he corr ect
rates of D R BF i n 2 3 5 10 20 di mensional case
re spect iv ely 
2 C i r cl e
We suppose t hat al l of t he sampl es of cl as s A are well 
dist ri but ed i n a circle T his ci rcl e l i es i n t he pl ane and it s
cent er i s t h e or igi na l poi nt and i t s r adi us is 1 T hi s ci rcle is
denoted LS c (0So

p( ) =
X ff C (1)
pY( I )px ( )d
( )

f( z 
)
f l 81
f l 91
F or di fi er ent k t he cor rect r ecogni t i on r at es of HSN net 
work and RBF network are calculated following Eq(1 91and
ar e shown in F ig3 I t can be seen t hat wi t h t he s iz e of sanl 
pl e set decreas i n g t he co rr ect r at es of H SN net wor ks i s kept
at a hi gh l evel whi le t he correct rat es of R B F net works drop
down qu ick ly T he r esu lt s also s upport t h at t he gen er al iz ati o
abi l it y of HSN network is st r onger t han t hat of R B F network 
esp eci al ly for sm al l sampl e set 
1
0 8
06
0 4
0 2
Sampl e set si ze
F ig 3T he corr ect rates of H SN networ ks and R B F network s
fo r different samp le set size T he cor rect r at es of H SN
net wor k is kept at about 99 ft he top l ine) when t he
sample count drops down and the di mension of featur e
space v ar ies How ever  the cor rect rates of R B F net
wor k drop down quick ly w hen the size of sample set
decreases (from up to down the curves indicate 2 3
510 and 20 di mensional cas e respecti vel y)
3 G e n er al di s t r i bu t io n
T he R B F n etwork s onl y cover t h e r egi on near t h e sam-
pi e po int s  w hi le t he H SN net wor ks can cover not on ly t he
same region of R B F net wor ks but al so t he r egi on n ear t he
li nes wh ich conn ect each cou pl e of sampl e point s  So t he HSN
net works can cover a l arger regi on t han R B F net works wi t h
th e same par amet er s especi all y w hen t he dist an ce between
sampl es is l ar ge T her efor e i n gen er al cas e t he H SN n et wor ks
can pr ovi de mu ch st r onger gener ali z at i on abi li ty t han t he R B F
net works 
V  A pp li c at i ons of B i omi met i c
P at t er n R ecogn i t io n
B i omimet ic pat t er n r ecogni t ion has been u sed i n many
< 
 ᅣ B R
ﺬ http://www.cqvip.com

376 Chinese J ournal of Electronics 2004
app l icat i ons s uccess fu l ly I n t hese ap pl i cat ions  B P R has
acqu ir ed b et t er r esul ts t han tr ad it i onal patt er n r ecogn it i on
met hods such as SV M and K NN met hod
I Iab l e 1 T he r esul t s of R B F SV M an d B P R
Tr aini ng R B F SV M B P R
samp les SV C
orrec t rate H SN correct rate
c o u n t
338 2598 9972 338 9987
25l 1925 9928 25l 9987
2l 6 1646 9456 2l 6 9941
192 1483 8838 192 98 98
182 1378 8095 182 9822
169 1307 7828 169 9822
T he fi r st app li cat i on of B iomi met i c patt er n r ecogn i t ion is
a r ecogn it i on syst em of omn i dir ect i onal ly or ien ted r igi d ob
jects on the horizontal surface_l J Ignoring the disturbance
t he di st r ibut ion r egi on of a cer t ai n cl ass i s t opologi cal ly homo
mor phi caI t o a ci r cl e So H SN net wor ks ar e used t o const r uct
t he cover in g set s of diff er ent cl asses T he SV M met hod w it h
R B F ker nel i s al so u sed as contr ol exper i ment  T he samples
for t ra in in g and t est ar e d i vid ed i nt o t hr ee samp le set s  T he
fi rst one contains 3200 samples of 8 obj ects (1ionrhinoceros
tiger dogtankbuscarand pumper )while the second one
cont ain s anot her 3200 sampl es t hat ar e col lect ed I at er fr om
the same 8 obj ects A third onewhich comprise 2400 sam
pies of another 6 objects (cat pug zebralittle lionpolar
bear and el ephant1is used fo r the false acceptance test  Al l
t he samples ar e mapped i nt o a 256dimensi onal feat ur e space 
T he H SN net wor ks ar e const r uct ed accordi ng t o t he t r ain i ng
s ampl es w hich ar e sel ect ed fr om t h e fi rst sampl e set  U n der
t he con di t i on t hat n o one sampl e in t he fi rst and second set
is mi scl assi fi ed and no on e in t he t hir d set i s accept ed fals el y
t he cor r ect r ecogn it i on r at es of B P R an d R BF SV M wi t h dif-
fer ent t r ai ni ng sets ar e shown n T ab le 1 I t can b e seen t hat
t he r esu lt s of B P R ar e much bet t er t han t he r esul t s o f SV M
met hod es peci al ly i n t he exp er i ment s w it h sm all samp le set 
T h is r esult is consonant wit h t he r esul t s of t h eor et i c anal ysi s
ver y wel 1
T abl e 2  T he r es u lt s f or O R L f ace dat a baNe
I Iest A I Iest B  A l
gorthms (245 pictures) (50 pictures)
SV M I C orr ect 190 F al se 3
(3 degree recogntlon 776 acceptance 6
poly nomial M iscl assi  2 C orr ect 47
ker nel 1 fi cati on 08 rej ecti on 94
SV M I I C or rect 193 F alse 5
fRB F recognti on 788 acceptance l O
M i scl assi  4 C orr ect 45 k
ernel 1 fi
cat ion 16 rej ect ion 90
C or rect 194 F alse 0
recognti on 792 acceptance O B P
R M i
scl assi  0 C or rect 50
fi cat ion O rej ect ion i 00
A n ot her a ppl i cat ion of B P R is a face r ecogn it i on syst em[

I f t he chan ges of face appear ance ar e consi der ed as dist ur 
bance t he dist r i but ion r egion is topologi cal ly homomor phi cal
t o an ar c when he t ur ns hi s face hor i zont al ly So t he HSN n et 
wor k i s ver y fi t t o const r uct t he cover i ng set  Ni net yone face
pi ct ur es of 3 person s ar e used t o const r uct t hr ee H SN net wor ks 
and 226 face pi ct ur es wer e u sed t o t est t h e cor rect r ecogn it i on
r at e of t h e sam e cl ass wh il e 728 p i ct ur es wer e used t o t est t h e
rej ection rate of the other cl assesT he correct recogni tion rat e
of the same class reaches 97 while the rej ect ion rate of the
ot her clas ses is 997  A s t he cont ras t t he correct recognit i on
rat e of the same clas s reaches 8982 while the rej ection rat e
of the other cl as ses s 9794 in K NN methodt6J
0 n 0 R L face dat abas ea famous dat abas e on w hi ch many
pat t er n r ecogni t ion met hods h ave been appl ied we acqu ir ed
a bet t er r esul t by B P R met ho d t han SV M met hods t  T h e
r esu lt s ar e s hown i n T abl e 2 I t can be seen t hat t he err or
rates fboth misclassifi cation rate and false acceptance rate)
r each 0 and t h e cor r ect r ate i mpr oves at t he same t i me i n
BP R compar ed t o SV M met hods 
V I  C on cl u sio n
Fr om above we can concl ud e t hat 
f1) BPR method emphas izes analyzing t he distribution of
a cer t ain cl ass sam pl es i n featu r e space fi r st l y
(2) The prior information on the distribution of the sample
set can impr ove t h e gen er al i zat ion abi li ty gr eat ly
f3) BPR has obtained better r esults than traditional pat
t er n r ecogn it i on met h ods  such as SV M and K N N i n many
ap pl i cat ions 
(4) The Hyper Sausage Neurons act better than RBF neu
r ons for cover in g t h e di str ib ut i on r eg ion of a cer t ai n cl as s i n
feat ur e sp ace T he comp ar i sons in t wo sp ecial cas es b et ween
H SN an d R B F sup por t t hi s ver y wel1
R ef er en ces
[1] wANG Shoujue Biomimetic pattern recognition (in Chi
nese) Acta Electroni ca Si nicaVo130 No10PP14171420
0 ct  2002
[2] WANG ShoujueLI ZhaozhouCHEN Xiangdong and WANG
Bai nan D iscussion on the basic mathemat ic model of neurons
in general purpose neurocomputer (in Chinese) Acta E lec
tronic a Si ni cav o129No5P P577- 580M av 2001
3] WANG Shouj ue and WANG Bai nan A nal ysis and t heory of
hi ghd imension space geometr y for art ifi ci al neural networks
fin Chinese) Acta El ectroni ca Si ni caVo130No1J an 2002
[4] WANG Shoujue Biomimetics pattern recognition Neural
Networks Soci ety fINNS ENNS J NNS) Newsletter o1
No1PP3 5M ar  2003
[5] wANG Shouj ueXU J ianWANG Xianbao and QI N Hong
M ul ti camera H umanf ace personal i dent ifi cati on system based
on the biomi metic pattern recogni ti on(i n Chinese)Acta Elec
troni ca S ini ca )13ON o1J an 2002
[6] WANG ZhihaiZHAO Zhanqiang and WANG Shouj ue A
met hod of bi omimetic pat ter n recogni t ion for face rec ogni
tionfin Chinese)Pattern Recognize Art cial I ntellegence
V ol _16No4D ec 2003
[7] WANG Shoujue and QU YanfengORL face databas e on the
biomimetic pattern recognition(in Chinese)accepted by Acta
E l ec troni ca Si ni ca
8] RA F i sherContri buti ons t0 Mathemati cal 0scsJ Wiley
New )rk  1952
[9] chen J igaotrans1ated by Qiu BingzhangQiu Hu 0 s c
P 0tt er n Rec0gﲸi ti 0ﲸpublished by Be ing I nst it ute of P osts and
r eleco mm unicat io ns l 989
ﺬ http://www.cqvip.com

B i omi met i c Pat t ern Recognit i on T heory and I ts A ppl icat ions 377
[1o VNVapnik and Chervonenkis AJ aTheory of Pattern Reco9
ni ti on NaukaM oscow  1974
[1l  BBoserIGuyon and VNVapnikA training algorithm for
optimal margi n classifi ersFifth Annual Workshop on Compu
tati on al L ear ni ng T heory PP144152P i tt sbur gh A C M 1992
[12Vladimir NVapnikThe Nature of Statistic Learn ing Theory
Spr inger  1995
131 A D Al eksandrov et a1 Mathemati cs I ts Essence M ethods
and R ol eV l013P ubli shers of t he U SSR A cademy of Sciences
M OSCOW 1956
[14Ryszard EngelkingDimension TheoryPWN-Polish Scient c
P ubl isher sW ar szaw a1978
_ WANG Shoujue specialist on semiconductor electronicswas born in June 1926 in ShanghaiHe graduated from the Department of Electrical Engineering TongJi University in 1949Since gradu ationhe served as Research Assistant in the Institute of RadiumPeking Academy in ShanghaiHe joined the Institute of Ap- plied PhysicsChinese Academy of Sciences
i n Beij ing i n 1956 and served as di rect or
of t he Semic onductor D ev ices L aborat or y si nce t he f oundi ng of t he
Semicond uct or I nst it ut eC A S i n 1960 He became deputy di rect or
and di rect or of t he Semi conductor I nst it ut e C A S in 1977 and 1983
r espect iv ely  H e was elect ed t he A cademici an of C hi nese A cademy
of Sc iences in 1980H e i s now a r esearch f el low in t he Semi conduct or
I nst it ut e C A S in charge of t he research l aborat ory on the Semi 
conduct or A r ti fi cial N eural Netw or k s H e i s al so a gu est P rof essor
at t he TongJ i University i n Shanghai and t he Zhej iang Uni versi ty
of T echnology in H angz hou P r ofessor Wang devel oped t he fi r st do_
mest ic made hi gh fr equency swit ching t ransistor in 1958fo r using
in t he ear ly high speed t ransi st or ized computer fo r comput ing i n
research wor ks on t he nuc lear Phy si cs He dev eloped also t he earl iest
si lic on planar tr ansistors and soli d state cir cuit s i n C hina i n 1963
and 1965t he P att er n Generat or fo r L SI mas k maki ng i n 1971 By
using hi s i nventi on on the I nt egrat ed High Speed F uzzy L ogic C i r
c ui t DY L  publ ished i n 1978he reduced the convert ing t ime of 8
bit DA convert or chi ps fo rm 80 as t o 4 as I n t he las t decade of
t he twent iet h centur y,he w or ked and got signi fi cant ac hievement s
on t he Semiconduct or A rti fi cial N eural N et work si ncludi ng chips
hardware mat hemat i cal model  al gori thm and appl icat i on on pat 
t ern r ecognit ion as w ell as pr ocess opt imizati on  P r of essor VCang
has r ec eiv ed t he fi rsclas s award of N at ional New P roducts P ri ze in
1964 the fi rstc lass as wel l as the sec ondcl as s award of Scient ifi 
cal and Technol ogical A chievement P r ize fr om C A S i n 1980 1983
1992and 1996t he fi r st clas s awar d of advance Sc ience and Tech
nology Prize Ber ing i n 2001 He also has twi ce recei ved the Na
t ional I nventi on P r ize in 1964 and agai n in 1996 H e has been given
the A ward of P rogr ess in Science and Technology i n 2001 (E mai l
wsj ue@redsemi acca)
Z H A O X i n gt ao was bor n in 1975
He recei ved his B S  degree i n Shandong
University i n 1997 maj ored i n E lect rical
E ngi neer ing A ft er graduat ionhe bec ame
a f acult y i n the D epart ment of E lectr i
cal E ngi neer ing Suzhou U niversity  F r om
1999 t o now  he entered t he I nst it ut e of
Semi conduct or s C hi nese A c ademy of Sci 
encesfi rst as a M S t hen as a P hD can
did ate Super v ised by A cademici an W ang
Shouj uehe is st udying the new algorit hms of art ifi ci al neural net 
wor k and t he appl icati ons of A N N in image processi ng and r ecog
niti o n now 
ﺬ http://www.cqvip.com