Federal gosudarstvennoe uchebnoe predpriyatie Chair of the System of Artificial Intelligence

b
a
61
1
directed
1
b - group of primeers with protivorechiem, class i - sostoyanie, in which right
4
opredelyaetsya primer i, arrows shown napravleniya vector gradient function
a - for single primeers with different answers;
nachalnoe
3 sostoyanie
Rice. 14. Graficheskoe predstavlenie protivorechiy v dannyx:
ie
2
2
Machine Translated by Google

---

DL_
2
CLASS
1
3.2
3.2
3
2.2
5.1
6.0
1.5
1.8
3
2
1.3
4.5
1.5
5.1
2.5
2.2
SH_
2
5.6
2.3
3
2.2
2.8
2.3
1.5
3
6.2
5.7
1
4.8
CHASH
1.5
1
1.4
3.2
5.0
1.5
2.8
2.8
4.0
2
5.6
Table 2.2. - Primer polucheniya protivorechivyx naborov vxodnyx
6.1
2.0
CLASS
1.8
3
3
2
2.2
2.6
1.3
4.9
1.5
1
2.8
2.8
1.3
2
2
3.1
2.5
2.6
dannyx pri sokrashchenii dliny opisaniya (chisla poley) on primere irisov
2.8
1.5
3
6.1
CHASH
3
5.5
1.5
SH_
2.0
Sleduet pomnit, chto vse algorithms obucheniya setey metodom
1.8
4.8
4.5
2.8
CHASH
2
2.3
4.6
2
6.9
2.2
1.8
1.1
1.3
3
3
6.5
5.6
Fishera
3.9
1.8
obratnogo rasprostraneniya oshibki (obshcheprinyatyy method, ispolzuemÿy
3
3.2
4.0
LEP
LEP
7.2
3.1
3.2
2
1.5
pri obuchenii s uchitelem) opirayutsya na sposobnost seti vÿchislyat
1.8
6.2
6.3
DL_
2.8
2
1.4
1.1
5.7
1.5
2.8
2.8
3
3
1.5
3.2
2
SH_
5.9
gradient funktsii oshibki po obuchayushchim parametram. Takim obrazom, obuchenie sostoit v
vychislenii gradienta i modifikatsii parametrov seti. Tem ne menee, sushchestvuet mnojestvo ne
gradientnyx metodov obucheniya, naprimer, metod pokoordinatnogo spuska, sluchaynogo poiska
i ryad metodov
2.2
1.4
1.3
2.8
LEP
SH_
2
6.0
0.2
2.3
2
0.2
1.3
2.8
4.6
3
62
Machine Translated by Google

---

tochki zreniya - when nalichii obuchennoy neyronnoy seti or drugogo
obucheniya seti. In practice the quality of training schitaetsya seti
rekomenduetsya ne ispolzovat pri statisticheskom analize). Obnarujenie
sostavlyaet bolee 50%. Obosnovat eto znachenie mojno so statisticheskoy
obsujdaetsya izbytochnost lineynoy svyazi (lineyno svyazannÿe parameters
dayut resheniya na dannyx s mnojestvom ÿÿÿÿÿÿÿÿÿÿÿ (ÿÿ ÿÿÿÿÿÿ ÿ
mojem s opredelennoy dostovernostyu pravilno predskazat vse chto
neural network. V etom sluchae obÿchno ne nablyudaetsya 100% pravilno
ustroystva, kotoroe budet pravilno reshat zadachu chashche, chem oshibatsya, my
Monte Carlo. This method is usually effective for "good" (neprotivorechivyx,
nepustyx) dannyx, chem gradientnÿe. S drugoy storony, napravlennÿe ili
gradientnÿe metody v chistom vide inogda voobshche ne
tesnoy nelineynoy svyazi - odna iz traditsionnyx zadach, gde primenyayut
brosaniya monetki). Sootvetstvenno, chem vÿshe etot pokazatel, tem luchshe
primerov testovoy vyborki) pokazÿvaet, naskolko «xoroshi» dannÿe - polnÿ,
neprotivorechivy, tochny i t. d. Odnovremenno poluchaem kachestvo
obuchilas set.
mezhdu bolshinstvom parameters nablyudaetsya nelineynaya svyaz (naibolee
lokalnymi minimumami, «yamami»).
Some of the most frequently used methods of training
are: 1) negradient methods: the method of sluchaynoy
arrow (one of the methods of Monte Carlo); pokoordinatnogo
method of descent (pseudogradientnyy); sluchaynogo search
method; Neldera-Mida method; 2) gradient methods: the method
of low-speed descent; kParTan; quasinyutonovskie methods
(BFGS). 2. Forecasting. Dlya polucheniya prognoza nujno
obuchit
neyronnuyu set po obuchayushchey vyborke - podmnojestvu polnoy vÿborki,
testirovat set po testovoy vyborke - podmnojestvu polnoy vÿborki. Quantity of
correct answers by test selection (percentage of test results
ugodno luchshe, chem pri ispolzovanii sluchaynyx metodov poiska (tipa
opredelennyx primerov. For example, for the task of immune plasmapheresis
udovletvoritelnÿm, esli protsent pravilno opredelennyx primerov
ÿÿÿ lyubym, vplot do lineynoy zavisimosti. On practice obÿchno
tesnaya mezhdu parametrami, sostavlyayushchimi leykotsitarnuyu formula), chto
Esli protsent pravilno opredelennyx primerov sostavil 100%, sleduet proverit
nalichie vo vxodnyx polyax parametra (parametrov), silno svyazannyx s polem
otveta. Character ÿÿÿÿ ÿÿÿÿÿ ÿ ÿÿÿÿÿ sluchae ÿÿÿÿÿ
63
Machine Translated by Google

---

ispolzuyut obuchenie po chetnÿm primeram, testirovanie po nechetnÿm i
Sushchestvuet vtoraya prichina nalichiya 100% pravilno opredelennyx
primerov pri testirovanii nablyudaetsya, naprimer, esli TV = OV (vyborki
ispolzuemoy vyborkoy. Set skoree «zazubrivaet» primer obuchayushchey
pri ispolzovanii obuchayushchey i testovoy vÿborok, poluchennyx na osnove
poslednem sluchae testirovanie mojet ne dat 100% pravilno
sushchestvuet obshcheprinyatyx apriornyx metodov opredeleniya isxodnoy
Frequency testing is divided into 2 tasks: testing for dannyx s
znacheniya otvetov in obuchayushchey vÿborke. Kak sledstvie - snijenie
testovoy) or posledovatelnoe doobuchenie (pri obuchenii ispolzuyutsya dve
lyubom
sluchae. ÿÿÿÿÿÿ pereobucheniya (overfitting) mojno predpolojit, esli kachestvo
seti; testirovanie dlya dannyx bez otveta (validation) pozvolyaet poluchit
ispolzovaniya vyborok v terminax teorii mnojestv. Empty OV - obuchayushchaya
opredelennyx primerov menshe 50%). At this nablyudaetsya low
In izbejanie effekta pereobucheniya dlya malyx vÿborok primenyaetsya
effekt obÿchno svyazan s izbytochnoy strukturoy seti po sravneniyu s
eksperimentami s ispolzovaniem neyronnyx setey s uchitelem.
V), or TVÿV, OVÿV, TVÿOV = V. The effect is 100% correctly defined
vyborki, po ocheredi. Dlya vÿborok srednego razmera i bolshix chasto
primerov: testirovanie po toy je vyborke, po kotoroy obuchali set, libo
odinakovye, t. e. soderjat odinakovye zapisi) or TV ÿ OV ÿ ÿ. V
vyborki, chem uchitsya po nim. In practice it sometimes happens, t. k. ne
naoborot. Pri ispolzovanii neyronnyx setey vÿborka schitaetsya maloy, esli
kolichestvo primerov imeet poryadok kolichestva poley ili menshe.
struktury neyronnoy seti, trebuemoy dlya resheniya konkretnoy zadachi. Inogda
pereobuchenie mojet byt rezultatom obucheniya s malym (nulevÿm) znacheniem
dopustimogo otkloneniya pri otsenke pravilnosti primerov. Poslednee privodit k
prodoljitelnoy «podgonke» otvetov seti pod tochnÿe
obshchey sovokupnosti, obuchenie po odnoy iz vÿborok (obuchayushchey or
opredelennyx primerov, no metodologicheski oshibka zdes prisutstvuet v
izvestnÿm otvetom (test) pozvolyaet otsenit kachestvo obuchennoy neyronnoy
prognoz libo zapolnit propusk v dannyx. Poslednee (dannye bez otveta)
vyborki), testirovanie po obshchey vyborke. Record the described options
testirovaniya seti nizkoe ili ochen nizkoe (protsent pravilno
pravilnosti otvetov on generalnoy sovokupnosti.
vyborka, TV - testovor vyborka, V - obshchaya vyborka (TV i OV poluchenÿ iz
podtverjdaetsya dannymi korrelyatsionnogo analiza i mnogochislennymi
oshibka obucheniya i vysokaya oshibka generalizatsii (obobshcheniya). Danny
«Skolzyashchiy kontrol» - training on all primaries without odnogo, testirovanie
on odnomu, where testovym primerom slujit odin primer
64
Machine Translated by Google

---

methodology. Sleduet otmetit, chto rangi, v otlichie ot srednix znachimostey
Significance of the input signal or parameter - relative rating
naoborot). Budem uporyadochenie chisel (znachimostey) proizvodit po
parameters uporyadochivaem po ubyvaniyu dlya kajdoy seti. Rasstavlyaem rang
rasporyajenii chislovÿe dannÿe (naprimer, znacheniya elementov vÿborki) nosyat v
toy ili inoy mere uslovnÿy character. Analysis takix dannyx trebuet
vyxodnyx signalov seti v sravnenii s analogichnymi pokazatelyami
znachimostey (with uchetom okrugleniya) have odinakovÿe color. Then for
srednim znachimostyam parameters (soxranyaetsya poryadok sledovaniya
rasschityvaetsya kak summa rangov kajdoy seti. Uporyadochivaem parameters v
series of experiments.
In order to choose the most important parameters, we use ranjirovanie
statisticheskix metodov (naprimer, predpolojenie o normalnosti zakona
znachenie mojet bÿt nebolshim (znachimosti usrednyayutsya po vsem setyam),
nablyudeniya nazÿvayut tot nomer, kotoryy poluchit eto nablyudenie v
otkazatsya ot analiza konkretnyx znacheniy dannyx, a issledovat tolko
serii neyrosetey prostavlyaetsya in sootvetstvii s vozrastaniem summarnogo
opredeleniya rangov po serii neyrosetey: znacheniya znachimostey vxodnyx
parameters, zavisyashchix drug ot druga i stepeni ÿtoy zavisimosti.
opredelennomu pravilu (naprimer, ot menshix znacheniy k bolshim ili
vneshney struktury (vxodnyx parameters) mojno ispolzovat analogichnuyu
vajnosti dannogo signalala or parameter seti dlya polucheniya zadannyx
velichine - at bolshix k menshim. Color primenyayut, esli imeyushchie c
dlya kajdogo znacheniya, nachinaya s maksimalnogo. Odinakovye znacheniya
parameters, bolee adequately used when analyzing vliyaniya parameters. V
bolshinstve sluchaev raspredelenie parametrov po rangam sootvetstvuet
kajdogo parameter opredelyaem summarnÿy rang dlya serii setey, kotoryy
znachimosti dlya drugix signalov, poluchennyx odnovremenno ili v sxodnoy
osoboy ostorojnosti, poskolku many predposylki klassicheskix
parameters by rank and by znachimostyam), but sometimes otlichayutsya. Srednee
znachimostey on a series of neurons. From the point of view of statisticians
raspredeleniya) dlya nix ne vÿpolnyayutsya. V etom sluchae imeet smysl voobshche
sootvetstvii s poryadkom vozrastaniya summarnogo ranga. Itogovyy rang po
uporyadochennoy sovokupnosti vsex dannyx - posle ix uporyadocheniya po
mojno nazvat resheniem zadachi prognozirovaniya. The term "testing" is used as
jargon, characteristic of the group NeuroComp. 3. Definition of values of input
parameters. Significance analysis - determination of relatively informative
parameters, definition
informatsiyu ob ix vzaimnoy uporyadochennosti.
Mojno predlojit ispolzovat sleduyushchuyu metodiku dlya
color. When determining the characteristics and colors after contrast
65
Machine Translated by Google

---

ix modules, by the sum of the square modules in t.
p. Otmetim, chto dannÿe pokazateli, kak pravilo, net smÿsla opredelyat, esli
pri testirovanii set pokazÿvaet menee 50% pravilno reshennyx
seti.
kontrastirovaniya posle etogo prodoljaetsya i izlishnie svyazi sokrashchayutsya.
Reshenie zadachi kontrastirovaniya pozvolyaet poluchit razrejennuyu,
logicheski prozrachnuyu neyronnuyu set, s bolee prostoy strukturoy, ponyatnoy
polzovatelyu. Dalee mojno pereyti k polucheniyu znaniy iz
prixodilo ne bolee n signals.
Dannye operatsii mogut privesti k sokrashcheniyu tselogo sloya neyronov
kontrastirovaniya s pomoshchyu bolshinstva vidov neyronnyx setey, v tom
otsutstvuet. Tem ne menee vozmojno konstruirovanie bolee tonkix nastroek
Kontrastirovanie mojet vklyuchat v sebya sleduyushchie operatsii:
sokrashchenie chisla vxodnyx parameters; sokrashchenie chisla
neuronov seti; sokrashchenie chisla synapsov seti; sokrashchenie
chisla neodnorodnyx (porogovyx) vkhodov neyronov seti; ravnomernoe
uproshchenie seti tak, chtoby na kajdyy neyron seti
tochnee opredelyaet poryadok sledovaniya parametrov po stepeni ix vliyaniya na
on minimizatsii vxodnyx parameters pozvolyaet opredelit nabor yadernyx
sloy ne mojet byt udalen). Realization of contrasting layers
4. Contrasting of neural networks (pruning, contrasting), minimization of the
number of input parameters. Conducting a series of experiments
tem ne menee, rang pokazÿvaet bolshuyu znachimost dannogo parameter. Rang
(v nekotoryx neyroimitatorax sushchestvuet ogranichenie, soglasno kotorogo
poluempiricheskix theory izuchaemogo yavleniya. Odnako dlya resheniya zadachi
seti, udalenie izbytochnyx svyazey, bez kotoryx reshaemaya zadacha prodoljaet
reshatsya tak je verno (s toy je tochnostyu). Frequent hearing
dlya neyronnyx setey s uchitelem: po summe proxodyashchix signalov, po summe
odnogo neyrona), posle udaleniya kotorogo sinapsy, prixodyashchie ranee na ego
vyxodnoe pole.
Sushchestvuet neskolko sposobov opredeleniya znachimostey parameters
parameters, kotorye ostayutsya vsegda ili pochti vsegda (v 80% setey). Ostalnÿe
parameters vÿrezayutsya in bolshinstve setey. Contrast - the process of minimization
of the number of neurons
predpolagaet nalichie minimalnogo kolichestva neuronov v sloe (sloy iz
dannyx - postroit poluempiricheskuyu teoriyu ili nabor
primerov, t. k. v etom sluchae vliyanie parametrov slaboe ili prakticheski
chisla vxodnyx signalov - ÿÿÿ sokrashchenie ix chisla do minimalnogo nabora, po
kotoromu mojno poluchit otvet na dannoy obuchayushchey vÿborke.
vxod, napravlyayutsya na vxod kajdogo iz neyronov sleduyushchego sloya. Process
kontrastirovaniya - minimization of the number of input signals. Minimization
66
Machine Translated by Google

---

dlya podachi seti, description nelineynyx funktsiy neyronov, funktsionirovanie neyronnoy
seti posloyno i poneyronno, pravila
ispolzuemyx arhitektur i metodov obucheniya neyronnyx setey opredelenie
proshche, chto svyazano s menshim chislom sinapsov i neyronov, t. e. mojet byt
postroena from sloistoy. Prichem realization dannogo metoda mojet bit
vÿpolnyayut posle obucheniya seti, prichem kontrastirovanie obÿchno trebuet
Sushchestvuyut algorithms and methods of training, which vÿpolnyayut
isklyuchitelno razrabotchikom. For example, in the neuroimitator NeuroPro v. 0.25
(razrabotchik V. G. Tsaregorodtsev) with contrasting sloistaya structure
v dalneyshem pri analize seti mojno imenovat eti signalaly v terminax
kajdyy sloy obyazatelno soderjit hotya by odin neyron. In
obshchem sluchae protsess postroeniya logicheski prozrachnoy seti (LPS)
svoditsya k uproshcheniyu seti, kotoroe obÿchno vklyuchaet v sebya: operatsii
synaptic ligaments.
ispolzovanie dvux tselevÿh funktsiy ili spetsialnoy tselevoy funktsii, kotoraya uslojnena
trebovaniem kontrastirovaniya. Only hardly soblyusti
s
prozrachnosti, est i drugie preimushchestva. Contrast accelerates
sushchestvenno bolee slojny, but menee effektyvny. Izvestno isklyuchenie - v
(binary synapses and neodnorodnye inputs in NeuroPro dalee ne
xode pervonachalnogo obucheniya bez primeneniya spetsialnyx algoritmov. 5.
Postroenie logicheski prozrachnoy neyronnoy seti, ispolzovaniem
kontrastirovaniya kak metoda. Prozrachnaya set mojet bit
bolee trudoemkie, chem obuchenie neyronnoy seti. For bolshinstva
neural network, t. k. kontrastirovannaya system realizuetsya sushchestvenno
perechislyayutsya ispolzuemÿe polya file dannyx, pravila ix predobrabotki
znachimosti i kontrastirovanie - spetsialnÿe operatsii, kotorye
menshe razmerom i deshevle.
sushchestvenno razlichnoy dlya raznyx neuroimitatorov i opredelyaetsya
normirovki vyhodnyx signalov seti v diapazon istinnyx znacheniy. Signalam,
generiruemÿm neuronami seti, prisvaivayutsya nekotorye imena i
soxranyaetsya, t. e. chislo sloev ostaetsya neizmennym (sloi ne udalyayutsya) i
mnogokratnogo doobucheniya i opredeleniya znachimosti vxodnyx parametrov i
kontrastirovanie in xode pervonachalnogo obucheniya. This oznachaet
subject area. Further, receiving verbalizovannoe description neyronnoy
In contrasting sets, chrome simple structure and logicheskoy
balance dvux procedure minimizatsii, poetomu takie algorithms and methods
kontrastirovaniya i binarizatsii vesov synapsov, neodnorodnyx vkhodov seti
srabatÿvanie obuchennoy neyronnoy seti v lyubyx neyronnyx sistemax. Krome togo,
ochen vajen vopros tehnicheskoy realizatsii obuchennoy
chisle sigmoidnyx, trebuetsya vÿpolnit dopolnitelnÿe vÿchisleniya, daje
monotonnyx neyronnyx setyax nablyudaetsya chastichnoe kontrastirovanie v
obuchayutsya). Then poluchayut verbalnoe description neyronnyx setey, where
67
Machine Translated by Google

---

neformalizovannoy task predskazaniya ili klassifikatsii.
liniyami peredachi informatsii. Odnako chasto ne udaetsya polnostyu
for the right solution zadachi nabor naibolee znachimyx vxodnyx
formulirovku criterion logicheskoy prozrachnosti - chem menshe neyronov v
otsenki priblijeniya k idealnomu logicheski prozrachnomu sostoyaniyu seti:
utverjdenie opiraetsya na tot fakt, chto chelovek mojet odnovremenno
likvidiruet vse izlishnie synapsÿ neyronnoy seti.
prixodyashchix na neyron signalov, my oblegchaem polzovatelyu zadachu
prozrachna. However, the number of layers is fixed when creating the set. Etot
kontrastirovaniya voznikayut otdelnÿe neyronÿ (or sloi neyronov), yavlyayushchiesya
prosto peredatchikami informatsii s predydushchego sloya na
deystvitelno znachimye promejutochnÿe signals (priznaki).
byt, pomogaem v imenovanii etogo priznaka. Only gelatelna
ochred iz seti doljny isklyuchatsya synapsy, a tolko potom - neodnorodnÿe vxody adaptivnyx summatorov.
otseivaet priznakovoe «shumovoe field», ostavlyaya minimalno neobxodimyy
resheniya zadachi - zapisat na estestvennom yazyke algorithm resheniya
i edinstvennyy vyxod). These neurons should be replaced at their own discretion
seti - na kajdyy neyron seti doljno prixodit ne bolee n signalov, gde n
Privedem kriterii, na osnove kotoryx mojno stroit nekotorye
izbavitsya ot izlishnix sloev i poetomu mojno vvesti okonchatelnuyu
priznakov.
4. Chem menshe chislo prixodyashchix na neyron signalov, tem luchshe. This
5. Chem menshe obshchee chislo sinapsov v seti, tem luchshe. Criterion
operirovat s dostatochno small chislom sushchnostey. Minimiziruya chislo
1. Chem menshe sloev neyronov v seti, tem set bolee logicheski
6. Synapse, according to which the signal is transmitted, logically neprozrachnee
svoego vesa na velichinu postupayushchego na sinaps signal. Poetomu v pervuyu
kriteriy vajen v tex situatsiyax, kogda chislo sloev seti izbytochno i posle
fakt ne trebuet poyasneniya - my minimiziruem chislo signalov, generiruemyx kajdym
sloem seti, chto pozvolyaet ostavit tolko
soderjatelnogo osmÿsleniya priznaka, generiruemogo neyronom, i, mojet
sleduyushchiy. Takie neyrony imeyut edinstvennyy vxod (i mogut imet takje
seti, mojno popÿtatsya vosstanovit pravila, sformirovannÿe setyu dlya
3. Chem menshe vxodnyx signalov seti, tem luchshe. This criterion
modification of this criterion and introduction of the criterion of uniformity