ÿ
ÿ
n
2
ÿ
ÿ
ÿ
0
1
ÿx ÿ
ÿ
x
Nelineynyy
.
preobrazovatel
summator
.
Tochka
vetvleniya
.
Vxodnoy
ÿ
ÿ
Rice. 6. Nonlinear signal transducer
Rice. 8. Formal neuron
Rice. 7. Tochka vetvleniya
ÿ
vnutrenniy rabochiy sloy
(hidden), ne poluchayushchiy vxodnyx signalov i ne
Perechislennÿe vÿshe prostye element neyronnyx setey mogut byt
generiruyushchiy vyxodnyx. In zarubejnyx istochnikax prinyato osobo
combined into more complex elements: block, layer, column. Chashche vsego
ispolzuetsya layer.
Layer layer
nabor neyronov ili summatorov, (psevodo)
odnovremenno vosprinimayushchiy vxodnuyu informatsiyu i
vÿdelyat neyronnÿe seti s dvumya skrytymi sloyami, po analogii s teoremoy
Kolmogorova o tochnom predstavlenii funktsiy, nesmotrya na otsutstvie
(psevodo) odnovremenno generiruyushchix vyxodnÿe signal.
Vÿdelyayut neskolko vidov sloev - vxodnoy, vyxodnoy,
skrÿtyy
pryamogo otnosheniya ÿtoy teoremy k neyronnym setyam. Chashche vsego
x
x
x
x
23
Machine Translated by Google
peredaet svoy vyxodnoy signal ostalnÿm neuronam, vklyuchaya samogo sebya.
Vyxodnymi signalami seti mogut byt vse ili nekotorye vyxodnÿe
From ukazannogo set of elements stroyatsya neyronnÿe
set bolshinstva
kak pravilo ne mogut byt realizovany bez bolshix trudnostey daje
users. Esli ne ogovoreno obratnoe, to kajdyy vyxodnoy signal i-go
nazÿvaemyx
dvoystvennyx setey. Seti takogo vida toje mojno razdelit na
analog or hybrid vychislitelnye mashiny. Poetomu asynchronous
mojet byt lyubym i nikak zaranee ne svyazano s kolichestvom neyronov v
vxodnÿe signals podayutsya to all neurons. Dlya
polnosvyaznoy seti vxodnoy summator neyrona fakticheski
information. Chashche vsego tak delayut po neznaniyu - skolko neyronov imet
dopolnitelnyx. Sloistye seti nepolnosvyazny.
These types are chastichno
rassmatrivat neyronnÿe seti, synchronno funktsioniruyushchie v discretnÿe
signal set, vtoroy - lineynuyu funktsiyu ot vyxodnyx signalov drugix
mnogosloynye.
V sloistyx setyax neyrony raspolojeny v neskolko sloev (risunok 9). Neyrony
pervogo sloya poluchayut vxodnÿe signalaly, preobrazuyut ix i cherez
neurons pervogo sloya poluchayut kajdyy vxodnoy signal. Osoboe
sloy i t.d. do k-go, which gives out output signals for the interpreter i
neyronnyx setey s 9 skrytymi sloyami.
tsifrovyx vychislitelnyx mashinax asynchronnye seti i neprerÿvnoe vremya
svoe naimenovanie: pervÿy - vxodnoy (Input), vtoroy - skrytyy (Hidden), tretiy - vyxodnoy
(Output). Polnosvyaznÿe
set imeyut odin sloy neurons; kajdyy neuron
razlichnyx neyrosetevyx architecture, sredi kotoryx ochen mnogo tak
priblijenno. For realizatsii asynchronous sets sleduet ispolzovat
sloya podaetsya na vxod vsex neyronov i + 1-go. The number of neurons in each layer
signal neyronov posle neskolkix taktov funktsionirovaniya seti. All
drugix sloyax. Although chasto ispolzuyut odinakovoe number of neurons in layers, this
nikak ne obosnovÿvaetsya algorithms or teknologiey obrabotki
tipy. Chashche vsego po svyaznosti vÿdelyayut osobo dve bazovyx arxitektury -
sloistye
i
polnosvyaznÿe seti , i eshche neskolko osobenno vajnyx
seti dlya dalneyshego obsujdeniya ne predstavlyayut interesa. Dalee budet
raspadaetsya na dva: pervÿy vÿchislyaet lineynuyu funktsiyu ot vxodnyx
neurons, poluchennyx on predydushchem shage.
sovpadayut s deleniem tipov po kolichestvu sloev - odnosloynye i
moment of time.
v kajdom sloe. Obshcheprinyatyy sposob podachi vxodnyx signals: vse
Sleduet osobo ogovorit, chto pri programmnoy realizatsii na
ispolzuetsya ot 1 to 3 skrÿtyx sloev, hotya vstrechaetsya ispolzovanie
tochki vetvleniya peredayut neuronam vtorogo sloya. Next srabatÿvaet
vtoroy
rasprostranenie poluchili trexsloynÿe seti, v kotoryx kajdyy sloy imeet
24
Machine Translated by Google
rasprostranennogo (Fig. 8) vida ne yavlyaetsya obyazatelnÿm. Sloistaya or
neskolkix parameters, znacheniya kotoryx menyayutsya ot neyrona k neyronu,
to set nazÿvayut neodnorodnoy (heterogeneous). If the activation function has
a sigmoidal type, it can be set
Selection in the quality of elements of
the set of neurons naibolee
neobxodimosti. For dvoystvennyx neyronnyx setey obÿchno vÿbirayut
uchastvuyushchie in ney elements. Edinstvennoe jestkoe trebovanie,
predÿyavlyaemoe arxitekturoy k elementam seti, - eto sootvetstvie
polnosvyaznaya architecture nelagaet sushchestvennyx ogranicheniy na
Selection of elements for sloistyx and polnosvyaznyx sets - po
arhitekturoy) chislu ego vkhodov.
Esli polnosvyaznaya set funktsioniruet do polucheniya otveda zadannoe
nazÿvayut odnorodnoy (homogeneous). Esli je ÿ zavisit eshche ot odnogo ili
chislo taktov k, to ee mojno predstavit kak chastnÿy
sluchay k-sloynoy seti, vse
sloi kotoroy odinakovy i kajdyy iz nix sootvetstvuet taktu
neuron with adaptive neodnorodnym lineynym summatorom na vxode.
Funktsiya aktivatsii neuronov (characteristic function) ÿ - nelineynyy
preobrazovatel, preobrazuyushchiy vyxodnoy signal summatora, - mojet bÿt
odnoy i toy je dlya vseh neyronov seti. V etom sluchae set
razmernosti vector of input signals of the element (it opredelyaetsya
nazÿvatsya sigmoidnaya set.
funktsionirovaniya polnosvyaznoy set.
25
Machine Translated by Google
vÿdavat vyxodnÿe. Takie seti mogut to polucheniya otveta funktsionirovat
Do'stlaringiz bilan baham: 
