Federal gosudarstvennoe uchebnoe predpriyatie Chair of the System of Artificial Intelligence

technology
mojno razlojit na bolee prostye operatsii) grupp operatsiy 0 urovnya, naprimer, normirovanie,
predobrabotka i t.p .; 2) operatsii - operatsii, ispolzuemÿe dlya neyrosetevoy obrabotki
neyronnyx setey s uchitelem (see p. 2.1.1, 2.1.2): 1.
Predvaritelnÿe operatsii - schityvanie faylov dannyx,
operations;
1) podgruppy operatsiy - neelementarnÿe sostavlyayushchie (kotorye
2.2 Formalnÿy yazyk i grammatika opisaniya neyrosetevoy
sootvetstvii s opisaniem neyrosetevoy teknologii obrabotki dannyx dlya
operatsiy.)
Sleduet otmetit, chto predvaritelnÿe operatsii realizuyutsya obÿchno
realizatsii; 3)
realization of operations - rassmotrenie konkretnoy realizatsii ili
algorithm vÿpolneniya (na urovne component set) operation. This uroven
formalnyx yazykov i grammatik. Opisÿvaetsya technology ispolzovaniya
1.1. Obyazatelnÿe operatsii ÿ schityvanie faylov dannyx, predobrabotka. 1.2. Neobyazatelnye
operatsii - predstavlenie dannyx. 2. Basic - creation of a set, reading of a set, training. 3.
Vspomogatelnye - testing, uproshchenie struktury seti, opredelenie znachimostey vxodnyx
parametrov, verbalizatsiya seti, soxranenie proekta (seti). (Name «vspomogatelnye» dano
ne po ix
In this section predlagaetsya model neyrosetevoy teknologii
ispolzovaniya neyronnyx setey s uchitelem, postroennaya na osnove
dannyx, rassmatrivaemÿe in samom obshchem vide bez ucheta konkretnoy
predobrabotka, predstavlenie dannyx.
vÿpolnenie osnovnyx i ne obyazatelnÿm - lyuboy iz vspomogatelnyx
realizovannymi in ideal neuroimitator. When building
models, we use four levels of detail
razrabotchika neuroimitatora (drugoy podobnoy programmnoy sistemy).
steps of importance, and in place - in order of fulfillment - in technologies
predusmatrivaet uchastie kvalifitsirovannogo polzovatelya or
71
Machine Translated by Google

---

+
situatsiya realizatsii operatsii - jelaemoe, no ne vsegda deystvitelnoe).
processing data, for example, products of the series Excel Neural Package (Winnet, Kohonen map),
realizovannÿe as a set of add-ons for Microsoft Excel, additional package for working with neural
networks in statisticalheskom
Frequently
Nekotorye predvaritelnÿe operatsii, v chastnosti schityvaniya faylov
Predvaritelnÿe Osnovnÿe Vspomogatelnÿe
-
system
+ / -
dannyx mojet vÿpolnyatsya kak v predobrabotchikax, tak i v nekotoryx
Statisticheskie
zagruzki dannyx provodyat ix normirovanie.
+
Integrated
-
integrirovannÿe packages, realizuyushchie srazu neskolko podkhodov dlya
vypolneniya (idealnyy sluchay, t. e. zdes i dalee opisana idealnaya
+ / Frequently
produced Statistica.
Table 3.1. - Gruppy operatsiy po mestu vypolneniya
-
+
+ / -
dannyx predusmotreny vo vsex ispolzuemyx paketax (v predobrabotchikax, neyroimitatorax i
integrirovannyx sistemax), zapolnenie probelov v
Predobrabotchik
neuroimitator. Some neuroimitators (not all) are automatic
packages
Neuroimitator
Privedem schematichno gruppirovku ukazannyx operatsiy po mestu ix
podxodyashchix paketax. Sushchestvuet neskolko podobnyx programs: FAMaster (vÿpolnyaet operatsii
zapolneniya propushchennyx dannyx, razrabotchik A. A. Rossiev), PredMake (razrabotchik L.A.
Zhukov), predobrabotchik A. A. Baturo, LingStat (prednaznachen dlya predobrabotki tekstov, razrabotchik
L. A. Zhukov), Jointer (razrabotchik N. V. Reshetnikova). Imeyutsya i
-
Frequently
72
Machine Translated by Google

---

Table 3.2. - Description of some neuroimitators (executed)
operations)
Pro
-
+
-
dannyx
+
+
+
+
euron
-
Text
-
vxodnyx parameters
Text
+
+
+
vnutrenney
Obuchenie
+
-
Neuro
-
+
+
-
Read file
+
structures
+
+
-
Text
-
Filling
Testing on
+
+
-
+
+
-
-
+
-
dannyx
Text
+
+
+
+
Iterator
probelov
-
+
Binary
vyborke
-
Format vxodnyx
+
+
-
-
Text
-
Text
Verbalization
Testing on
+
+
+
+
(“-1”)
-
-
GA Neural
Text
Dbase /
-
-
+
+
-
-
Text
-
Plan-ner
Vyxodnyx format
primer
+
+
-
Normirovanie
+
-
-
Text
-
Text
Para
dox
+
Neuro
+
-
+
-
+
-
MultiN
-
+
+
dannyx
-
-
+
-
dannyx
-
Office
Contrasting
dannyx
+
+
+
+
+
+
Dbase,
+
Xranenie setey
-
Preobrazovanie
Text
+
+
+
Editing
-
-
vneshney structure
Evolu
tion
Neuro
+
+
poley
+
+
Text
+
Contrasting
-
Para-dox
Significance
73
Machine Translated by Google

---

In deystvitelnosti je gruppirovka operatsiy ne yavlyaetsya takoy
ochevidnoy i ne vse operatsii iz privedennyx grupp realizuyutsya v polnom
obÿeme. V bolshinstve sluchaev kajdyy avtor sootvetstvuyushchego
programmnogo paketa sam opredelyaet i traktuet neobxodimyy nabor operatsiy
dlya realizatsii. Poetomu razumnee v dannom kontekste privesti bolee
detalizovannuyu klassifikatsiyu kak po operatsiyam, tak i po otdelnym
programmnym produktam (tabl. 3.2).
In tselom osnovnÿe i nekotorÿe iz predvaritelnyx operatsiy yavlyayutsya
obyazatelnymi dlya vÿpolneniya, a vspomogatelnÿe yavlyayutsya
neobyazatelnymi (dopolnitelnymi) i mogut ne vÿpolnyatsya pri obrabotke
dannyx. V dannoy rabote rassmatrivayutsya vozmojnÿe posledovatelnosti
deystviy pri obrabotke dannyx, vÿpolnyaemÿe dlya odnoy neyroseti (odnogo
proekta). V kakom-to smÿsle zdes rassmatrivaetsya jiznennyy tsikl odnoy
neyronnoy seti (proekta) i rezultaty obrabotki dannyx s ee pomoshchyu. Iz
etogo sleduet, naprimer, chto sozdanie neyronnoy seti obyazatelno prisutstvuet
v slove rassmatrivaemogo formalnogo yazyka odin i rovno odin raz.
Analogichnym obrazom mojno rassmatrivat dannoe utverjdenie v terminax
ispolzovaniya odnogo proekta, esli operatsii nad proektom podrazumevat kak
operatsii, realizuemÿe nad kajdoy neyronnoy setyu, soderjashcheysya v
proekte. Pri etom ukazannÿe operatsii i ix poryadok yavlyaetsya odinakovÿm
dlya vsex setey proekta. Esli rassmatrivat jiznennyy tsikl obrabotki dannyx,
kotoryy yavlyaetsya bolee slojnÿm protsessom, opisanie formalnoy
grammatiki v etom sluchae budet otlichatsya ot razrabatÿvaemoy. At rasshirenii
polya operatsiy mogut prisutstvovat srazu neskolko setey, v obshchem sluchae
raznoy arxitektury i struktury, chto budet nakladÿvat opredelennÿe ogranicheniya
na yazyk (osobenno pri opisanii nijnego urovnya detalizatsii). Protsess
obrabotki dannyx potentsialno mojet bÿt beskonechen i nachinatsya ne s
samogo nachala i provoditsya ne do kontsa s tochki zreniya opisaniya odnoy
seti. Mojno vzyat soxranennuyu neyroset i s nekotoroy periodichnostyu provodit
razlichnÿe eksperimenty, ispolzuya opredelennÿe operatsii ili nabory operatsiy.
Or I will use soxranennuyu set na drugix dannyx dlya drugix zadach. Krome
togo, vozmojno sozdat ili prochitat neskolko neyronnyx setey i provodit otdelnÿe
(v obshchem sluchae razlichnyy nabor operatsiy) eksperimenty dlya kajdoy iz
setey. Analogichno protsess obrabotki dlya otdelnyx ili vsex setey mojet
provoditsya ne s nachala (seti mogut byt sozdany ranee) i ne do kontsa (seti
mogut byt ispolzovany dlya resheniya drugix zadach na drugix dannyx).
Sleduet otmetit, chto formalnoe opisanie neyrosetevoy teknologii (a
imenno alfavit yazyka opisaniya) mojet bÿt predstavleno dlya dvul razlichnyx
urovney detalizatsii: pervyy uroven vklyuchaet privedennye operatsii bez
rassmotreniya operat realizatsiyheski
74
Machine Translated by Google

---

seti
klassifikatsiey yazÿkov programmirovaniya: yazÿki vÿsokogo urovnya (analog
data processing. Predvaritelnÿe operatsii uslovno razdelim na
tablitsy, schityvaniya,
predobrabotki dannyx
details). Takim
obrazom, ves protsess neyrosetevoy obrabotki dannyx
and filling of spaces, normalization), operatsii schityvaniya faylov
B Predvaritelnye
neurons, definition
M Osnovnÿe
neyroimitatora neobxodimo rassmatrivat protsess raboty s odnoy setyu
obuchenie neyronnoy seti. Vspomogatelnÿe operatsii - eto testirovanie, uproshchenie
neyrosetey, opredelenie znachimostey vxodnyx parametrov, verbalizatsiya, zagruzka
seti (otkrytie file) i soxranenie seti v file. ÿÿÿÿÿÿÿÿ ÿÿÿÿ ÿÿÿÿÿ ÿÿÿÿÿÿ ÿÿÿ ÿ ÿÿÿÿÿ
ÿÿÿÿÿÿÿÿ ÿÿÿÿÿÿÿÿ, ÿÿÿ ÿ
parameters, verbalization,
soxranenie seti (proekta) 1
uroven
(projects). Sootvetstvenno, formalnoe description obobshchennoy obrabotki
We present the described technology in the form of a formal description. Imeem
sleduyushchiy neterminalnyy alphabet. Nije perechislenÿ simvolÿ 0 i 1 urovney:
A Vspomogatelnye
I Predstavlenie
Before main Operatsii sbora dannyx, sozdaniya
operatsiy uje na baze konkretnoy ÿlementnoy sostavlyayushchey seti (neuronov,
synapsov, vychislenie gradientov i t.p.). Here you can find an analogy with
mere dlya vsex variantov i vsex osobennostey.
Rassmotrim tri gruppy tehnologicheskix operatsiy pri neyrosetevoy
Main
operations
Data collection, creation of structures
pervogo urovnya detalizatsii) i nizkogo urovnya (vtoroy uroven
neskolko podgrupp: operatsii predstavleniya dannyx (sbor dannyx, struktura tablitsy
bazy dannyx, predvaritelnyy vÿbor polya otveta), predobrabotka (preobrazovanie
simvolnyx poley i poley tipa data, poisk
Creating / reading seti and obuchenie
Testing, uproshchenie
operations
rassmatrivaetsya s nekotoroy promejutochnoy tochki zreniya. For development
dannyx. Osnovnÿe operatsii predstavlyayut sozdanie / chtenie (zagruzka) i
znachimostey vxodnyx and vnutrennix
Auxiliary
operations
(project). Dlya obrabotki dannyx harakterna rabota s neskolkimi setyami
vspomogatelnyx.
operations
dannyx namnogo slojnee i edva li kogda-libo budet realizovano v polnoy
realizovanÿ v neyroimitatore), vtoroy - vklyuchaet realizatsiyu etix
0 uroven
operations
Data
75
Machine Translated by Google

---

Data
answer
nekotoromu algorithm
E Filling
BMA {BA},
nechislovyx poley, search i
P Predobrabotka Preprocessing
Empty data
Zapishem nabor dopustimyx vyrajeniy s ispolzovaniem
B M.
Test / Validate Testing on primer and on
Obuchenie
seti Uproshchenie struktury
seti, naprimer, operatsii
Learn
{BMA},
Normalization Preobrazovanie dannyx dlya
structure seti
Sozdanie struktury tablitsy,
predvaritelnyy vybor polya
BM {BMA},
N Normirovanie
tablitsy bazy dannyx, vybor polya
opredelennomu diapazonu
Fill in the gaps on
description of dannyx
BMA {MA},
poles and poles type data in drugix
BM {BA},
dannyx
probelov
answer
T Testing
filling of spaces,
normalization
L Obuchenie seti
neterminalnyx simvolov 0 urovnya (B, M, A):
{BM},
contrast and binaryization
C Uproshchenie
vyborke
privedeniya znacheniy k
introducing
Contrasting
D Predvaritelnoe
BMA {BM},
76
Machine Translated by Google

---

binary
vxodnyx
At urovne operatsiy polzovatelya rassmatrivayutsya konkretnÿe
signals so, mine on
parameters
importance
or
vnutrenney
vxodov s doobucheniem
opredelenii
synapses
contrast
uniform pruning Udalenie signalov tak, chtoby na neyron
prikhodilo
neodnorodnyx vxodov
vxodnyx parameters
neurons
Open the file
structure seti
{-1, 1}
(neuroimitator operatsii, predostavlennÿe konechnomu polzovatelyu dlya ispolzovaniya po svoemu
parameters with doobucheniem
2.2.1. Uroven operatsiy polzovatelya (2-y uroven)
pri
(contrast)
Description
neurosystem),
Udalenie neodnorodnyx
in Contrasting
with Contrasting
znachimostey
rassmatrivayutsya bez ucheta konkretnoy realizatsii, v samom obshchem vide), vxodyashchie v sostav
podgrupp. Predpolagaetsya nekotoraya ideal system
ne bolee N)
doobucheniem
znachimost
reading data
k znacheniyam iz mnojestva
verbalization
doobucheniem)
vyxodnogo
parameters, pokazÿvayushchix
inputs contrast Udalenie
and Opredelenie
contrast
uproshchenie seti
doobucheniem
terminal alphabet: Schityvanie file
neuron ix prixodilo
operatsii (predusmotrennÿe v idealnyx neyroimitatorax, operatsii
Udalenie sinapsov s
ix
dannyx
opisyvayutsya
Privedenie vesov synapsov
synapses
v Verbalization set
this Ravnomernoe
vxodnyx or vnutrennix
neurons contrast Udalenie neyronov s
usmotreniyu. On dannom urovne eti operatsii predstavlyayut soboy bukvÿ
field
ne bolee N signals (s
Raschet znachimostey
non-uniform
cs Contrast
cn Contrasting
b1 Binary set ± 1
rd
77
Machine Translated by Google

---

b2 Binary set ± 0.5, ± 1
b4 Binary set ± 0.1, ± 0.2, ±
0.3, ± 0.4, ± 0.5, ± 0.6,
± 0.7, ± 0.8, ± 0.9, ± 1
tv Test seti po
tp Testing on
rp ÿÿÿÿÿÿÿÿ (chtenie
b3 Binary set
tt Test seti po
mk ÿÿÿÿÿÿÿÿ ÿÿÿÿ
st
pri
± 0.25, ± 0.5, ± 0.75, ± 1
file) set
primer
Primary testing
d Sbor dannyx
Privedenie vesov synapsov
Privedenie vesov synapsov
primer at izvestnom
answer
file
vyborke pri
Creating structures
vyborke
{-1, -0.5, 0.5, 1}
{-1, -0.9, -0.8, -0.7, -0.6, -0.5,
-0.4, -0.3, -0.2, -0.1, 0.1, 0.2,
0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9,
1 }}
pri neizvestnom pole
Sozdanie seti (choice)
Determination of poley, ix
binary
answer
(project)
Save the set for
save
pri
dannyx
content
Testing
{-1, -0.75, -0.5, -0.25, 0.25, 0.5, 0.75,
1}
validate
make
binary
your Testing on
binary
s Save storage v
izvestnom znachenii polya
pri izvestnom pole otveta
data
k znacheniyam iz mnojestva
k znacheniyam iz mnojestva
tablitsy
test
Vibration testing
field answer
Testing
Privedenie vesov synapsov
Zagruzka soxranennoy seti
read project
neizvestnom pole
struktury, polya otveta, characteristic
neurons) Ruchnaya operation vvoda
k znacheniyam iz mnojestva
dalneyshego
format in smyslovogo
ispolzovaniya
neizvestnom pole answered
answer
table structure
78
Machine Translated by Google

---

normirovaniya
diapazon [a, b], gde a <0, b>
1, pryamaya otobrajaetsya v
get results
vyborochnoe srednee
nonlinear
normalization
radius
poley
structures in dannyx)
poley tablitsy dannyx
modify table
edinichnogo
at interval [0,1], e.g.
1, 2 or 3) at interval [0,1]
kusochno lineyno: na levuyu
ispolzovaniem
granitsu - everything, that's right
(predvaritelnyy)
Normirovanie na interval
normalization
arctg, logarifmicheskogo
(kajdyy vector delitsya na
tablitsy
Normirovanie on the sphere
linear
normalization
[0,1] nestrogo, not linear
interval [a, b]
numeric format i
[-ks, + ks], where s - est
kotorogo xotelos by
linear
will be
normirovanie
sferu edinichnogo
Change of structure
vvedenie dopolnitelnyx
s
linear
kvadratichnoe otklonenie, k -
kolighestvo sigm (obÿchno
pryamuyu, no interval [-ks, +
ks] otobrajaetsya in [0, 1]
Nelineynoe normirovanie
sphere
normalization
(change
granitsu otobrajaetsya vse,
chto levee, na pravuyu
structure
approximately
linear
normalization
sigmoidy, functional cos or
radius
normirovaniiya
[0,1] linearly
Normirovanie na interval
svoyu dlinu)
Preobrazovanie poley k
Normirovanie iz intervala
select target field Select parameter, for
f Preobrazovanie poley modify fields
[-ks, + ks], result
nl Normirovanie
on Normirovanie kusochno
nd Normirovanie na
sy Vybor polya otveta
mt Change structures
nn Nelineynoe
ns Normirovanie na
79
Machine Translated by Google

---

spuska
srednim znacheniem po
on the field
naiskoreyshego spuska
BFGS method
Primechanie: osnovoy dlya predlojennogo list of metodov obucheniya
zero fill
probelov
BFGS method
Training method Monte Carlo
maximum value
chisle razrabotannyx v g. Krasnoyarsk, dopustimo dopolnenie dannogo
minimal
probelov
Obuchenie
sluchaynogo poiska
Training with kParTan method
Bump
polyu
method
Obuchenie
naiskoreyshego spuska
Obuchenie
synapses
minimalnym znacheniem
Filling
probelov
maximum
pokoordinatnogo
kvazinyutonovskim
poslujili metody, ispolzuemÿe v neyroimitatorax Neurogenesis, STATISTICA Neural Networks, i ryade
drugix podobnyx programm, v tom
method
class, esli class izvesten
Filling
Monte Carlo
method
perechnya (strogo govorya, chislo modifikatsiy metodov obucheniya potentsialno
on the field
average fill
Filling
Obuchenie
pokoordinatnogo spuska
sluchaynogo search
kParTan
Vozmushchenie
Obuchenie
Neldera-Mida
srednim znacheniem po
class
probelov
Neldera-Mida
lc Obuchenie
class average fill Zapolnenie
Fill in the blanks
minimal fill
maximal fill
method
kvazinyutonovskim
vesov
lm Obuchenie metodom
ln Obuchenie metodom
lb Udar seti
ec Filling srednim po
ls Obuchenie metodom
lr Obuchenie metodom
lk Training method
lp Training method
em Zapolnenie
ex Fill
ez Zapolnenie nulem
ea Filling srednim
80
Machine Translated by Google

---

(1)
neyroimitatorax v vide nastroyki ili optsiy (chast 3 urovnya detalizatsii - realizatsii
operatsiy). V etom sluchae na urovne operatsiy polzovatelya
terminal alphabet: S :: =
{BM [A]} | {BM [A] {BA}} | {BMA {MA}}
A :: = i | C | v | s | T
(sootvetstvenno mojno ispolzovat tolko odin terminalnyy simvol dlya
(3)
Ukajem vse pravila dannoy grammatiki (na urovne operatsiy
Analogichnym obrazom mojno dopolnyat operatsii testirovaniya, binarizatsii i
kontrastirovaniya (naprimer, polezno vvesti operatsiyu
T :: = tt | tv | tu | tp
slozhnost grammatiki, uvelichivaetsya tolko chislo pravil. I represent
some operations (elements of the alphabet) in the form
L :: = lm L | lp L | lr L | ln L | ls L | lk L | lc L | lb L
S ÿ BMAK | BMK
S :: = in C | cs C | cn C | cu C | ce C | b1 C | b2 C | b3 C | b4 C
ispolzuemye method). Obÿchno vÿbor method obucheniya realizuetsya in
elements of the neterminalnogo alphabet, in vide sovokupnosti elements
K ÿ BMK | BMAK | BAK | BM | BMA | BA | ÿ
obuchenie yavlyaetsya odnoy operatsiey bez ukazaniya konkretnogo metoda
B :: = [I] rd
PM :: = mk L | rp L | rp
P :: = [f] [E] N
(2)
E :: = ez | ec | ea | em | ex
training set).
A :: = i A | CA | v A | s A | TA
contrast of sloev neurons). Completion of operations does not affect
L :: = lm | lp | lr | ln | ls | lk | lc | lb
users):
Bekusa-Naura, nachinaya s samogo verxnego urovnya i dalee detaliziruya
beskonechno, poetomu zdes privedeny nekotorÿe osnovnÿe, chashche
C :: = ci | cs | cn | cu | ce | b1 | b2 | b3 | b4
A ÿ AMA | AMAK | AM
I :: = [d] st [sy]
N :: = nl | na | nd | nn | ns | ns nl
81
Machine Translated by Google

---

(17)
C ÿ ÿ
I ÿ ÿ
(8)
neuroimitator, which predpolagaet rabotu with konkretnoy strukturoy i
M ÿ mk L | rp L | rp
A ÿ i | C | v | s | T
(19)
(10)
Primechanie: sbor dopolnitelnyx dannyx (dopolnenie
(12)
E ÿ ez | ec | ea | em | ex
ogranicheniy:
a) pervoy operatsiey doljno bÿt schityvanie file dannyx; b) operatsii
predobrabotki nelzya provodit do schityvaniya fayla
technologic operations, t. k. neuroimitator obchno ne predpolagaet
B ÿ rd P | I rd P
(5)
(14)
v) operatsii sbora dannyx nelzya provodit posle opredeleniya
(7)
I ÿ d D | d
(16)
(trivialnoe ogranichenie) - rassmatrivaetsya s tochki zreniya ispolzovaniya
A ÿ i A | CA | v A | s A | TA
naborom dannyx;
(9)
(18)
(11)
L ÿ lm L | lp L | lr L | ln L | ls L | lk L | lc L | lb L
Ukajem, chto pravila vyvoda vvedenÿ s uchetom sleduyushchix
sushchestvuyushchego file dannyx) predpolagaet obrÿv realizatsii dalneyshix
dannyx (trivialnoe ogranichenie);
(4)
(13)
postoyannogo obrashcheniya k dannyx file. In this sluchae trebuetsya zanovo
P ÿ mt f EN | mt f N | EN | N
C ÿ ci C | cs C | cn C | cu C | ce C | b1 C | b2 C | b3 C | b4 C
(6)
(15)
struktury bazy dannyx i predvaritelnogo vybora polya otveta
L ÿ lm | lp | lr | ln | ls | lk | lc | lb
N ÿ nl | na | nd | nn | ns | ns nl
T ÿ tt | tv | tu | tp
C ÿ ci | cs | cn | cu | ce | b1 | b2 | b3 | b4
D ÿ st sy | st
82
Machine Translated by Google

---

neyrosetevoy teknologii, imeet nekotorÿe osobennosti i vozmojnÿe
dopolnitelnyx dannyx or dopolnenie poley pozvolyaet bolee tochno
neyrosetevoy teknologii otdelnÿe svoystva grammatiki ne
operatsiy v samom neuroimitatore i chashche vsego ne imeet sredstv dlya ego
dannyx o korrektnoy posledovatelnosti operatsiy i ix bolee obshchey
results. g)
operatsii normirovaniya nelzya provodit do izmeneniya struktury i
operatsiy polzovatelya, poetomu realizatsiya tsiklichnosti zdes ne uchtena.
formalnyx yazykov i grammatik. According to the hierarchy of operations,
nalichiya takix tsiklov operatsiy. Perechen takix kriteriev traditsionen i
d) all operations pervonachalno (at pervom vxode) obyazatelno
ogranicheniy na kombinatsii operatsiy 0 urovnya.
Dannaya grammar does not imply the use of cycles at
e) osnovnÿe operatsii nelzya provodit do schityvaniya fayla dannyx
usloviy ili kriteriev vozniknoveniya dannogo sobytiya. Realization vhoda i
proisxodit po zadannoy vyborke do trebuemogo urovnya tochnosti (zadaetsya v
vnutrennego algorithm, which describes the level of realization
samogo nachala. This approach can be very effective, as is the collection
(trivialnoe ogranichenie); z)
poslednyaya operation should not be predvaritelnoy.
Utverjdeniya i ogranicheniya, kasayushchiesya dannoy grammatiki, privedeny
dalee. First of all, when developing a formal description
element, esli trebuemaya tochnost ne dostignuta, element ne vÿrezaetsya, perexod
k sleduyushchemu i t. d.). Dannaya grammar, rassmatrivaemaya in context opisaniya
opredelit vzaimosvyazi mejdu parameters i nayti obosnovanie
rassmatrivalis. Pravila vyvoda stroilis na osnove empiricheskix
corrections with users. Grammar is developed only up to the level
modifications, kotorye otlichayut ee ot traditsionnogo podkhoda v teorii
At this level of detailing is possible to describe the criteria
preobrazovaniya tipov poley i zapolneniya probelov;
gruppirovki. Otmetim, chto neterminalnyy simvol K vveden dlya ucheta
doljny nachinatsya s predvaritelnyx i osnovnyx;
vÿpolnenii posledovatelnosti operatsiy, chto podrazumevaet nalichie
obshcheprinyat dlya dannoy predmetnoy oblasti (neyroinformatiki): obuchenie
(trivialnoe ogranichenie); j)
vspomogatelnÿe operatsii nelzya provodit do predvaritelnyx
otkrÿt file dannyx s dopolnennymi zapisyami i provodit vse operatsii s
vyxoda iz tsikla predstavlyaet soboy organizatsiyu sootvetstvuyushchego
nastroykax pri sozdanii seti), sokrashchenie chisla elementov - posledovatelno s
doobucheniem (popÿtka obuchit set bez otdelnogo
83
Machine Translated by Google

---

2.2.2 Postroenie atributnyx grammatics for neyrosetevoy teknologii
oblasti.
Poetomu v dalneyshem primem pervonachalnye pravila grammatiki za
pozvolyaet vÿdelyat razlichnÿe vidy otdelnoy operatsii (naprimer, neskolko vidov
normirovaniya, obucheniya i t.p.). Dopolnitelno mojno
prisvaivaniya, so and sostavnÿe operators.
At this analogichno urovnyam tehnologii nujno uchest nalichie
simvoly (i terminalnye simvoly, esli rassmatrivat uroven realizatsii
tolko pravila, kotoryh kasayutsya izmeneniya ili modifikatsii.
chast pravila sostoit iz lokalnyx ob'yavleniy i semanticheskix deystviy. In
kachestve semanticheskix deystviy dopuskayutsya as atributnÿe
neyrosetevoy teknologii «razraslas» to semeystva grammatical, kotorye
neobxodimost dopolnitelnogo opisaniya nekotorÿh parameters, svyazannyx
nastroykoy or nastroykoy on umolchaniyu) and for professionals
For realizatsii grammar neyrosetevoy teknologii voznikaet
The grammar of the alphabet can be called 3-level: 0 - group operations, 1 -
subgroups, 2 - operations. Krome togo, iznachalno predlojennaya grammatika
opisaniya
urovney ili grupp dannyx parametrov po stepeni «slojnosti» neyroimitatora v
smysle realizatsii nekotoryx spetsificheskix funktsiy: uslovno dlya nespetsialistov
(realizatsiya osnovnyx funktsiy s minimalnoy
operatsii otkrytiya file dannyx. In this context neterminalnye
operatsiy) mojno rassmatrivat kak polu terminalnÿe ili chastichno
grammars (by analogy atributnÿe grammars arepolzuyutsya for
(usecheniem or rasshireniem) in sootvetstvii s analizom predmetnoy
description of semantic text programming).
Description atributnoy grammar sostoit iz ÿÿÿÿÿÿÿ ÿÿÿÿÿÿÿÿ
2.2.2.1. Level "non-specialist" (common level)
opisÿvayut razlichnÿe modifikatsii pervonachalnoy, poluchennyx ne stolko
ekvivalentnymi preobrazovaniyami, skolko izmeneniyami pravil vyvoda
with ix semantics. Dlya etogo sushchestvuyut tak nazÿvaemÿe atributnÿe
(ability to fine-tune the architecture of the set and its parameters).
vvesti parameters, opredelyayushchie, naprimer vid vÿborki, ee imya dlya
bazovÿe i pri opisanii pravil grammatiki semeystva budem ukazÿvat
attributes for each attribute grammar and type of attribute. Pravila sostoyat iz
sintaksicheskoy i semanticheskoy chasti. Semanticheskaya
In the description of the terminal alphabet entered indexes, chto
attributes and razdela rules. Razdel opisaniya attributes opredelyaet composition
84
Machine Translated by Google

---

rdt (source)
Nekotorye operatsii, naprimer, chteniya fayla dannyx ili zagruzki, soxraneniya seti
trebuyut nalichiya parametrov, kotorye odnoznachno
technology). For example, for operative reading file dannyx mojno
rdg (svod.dbf).
Sleduet otmetit, chto na dannom urovne v kachestve atributov
dannyx). Ispolzovanie parameters pozvolyaet tochnee opredelit context i
rdt (svod2.dbf)
neposredstvenno to neyrosetevoy teknologii, t.e. ne uchityvayutsya attributes,
otnosyashchiesya neposredstvenno k teknologii samogo protsessa chteniya: naprimer,
chtenie zagolovka, struktury bazy dannyx (zagolovki poley, tipy
rdl (source)
For example, operatsiyu chteniya fayla dannyx mojno predstavit
Strogo govorya, dlya dannogo opisaniya trebuetsya ispolzovat
traditsionnoy i ne trebuet dopolnitelnogo opisaniya v dannoy rabote.
reading data (learn sample)
vyvoda slov yazyka iz pravil grammatiki. Tolko v dannom sluchae s kajdym
rdl (svod1.dbf)
rdg (svod.dbf)
shagom proisxodit concretization (terminal symbol
on this level of detail of the description
predobrabotki (P), operatsii uproshcheniya - vo vspomogatelnÿe i t.p.
reading data (general sample)
vyborki dannyx iz fayla svod.dbf formalno mojet bÿt zapisan kak
opredelyayut dannuyu operatsiyu (naprimer, ukazanie konkretnogo file
operatsii chteniya dannyx ispolzuetsya ogranichennÿy set, otnosyashchiysya
predstavit sleduyushchiy “vyvod” (otdelno dlya kajdoy vÿborki):
dannyx
obshchiy smÿsl posledovatelnosti operatsiy.
dannyx) i t.p. Predpolagaetsya, chto realization process yavlyaetsya
rdg (source)
ÿ
sleduyushchimi terminalnymi simvolami:
schityvanie obuchayushchey vÿborki
svoeobraznÿy «vyvod», kotoryy analogichen podkhodu, primenyaemomu dlya
reading data (test sample)
terminalnye. For example, operatsii normirovaniya (N) vxodyat v operatsii
rassmatrivaemoy operatsii
In dannom sluchae characters, opisÿvayushchiy operatsiyu chteniya obshchey
ÿ
ÿ
ÿ
ÿ
ÿ
85
rdg schityvanie obshchey sovokupnosti
rdl
rdl
rdt
rdt schityvanie testovoy vyborki
rdg
Machine Translated by Google

---

g
V dalneyshem budem ispolzovat dve ravnoznachnÿe form zapisi
attributes. Predpolagaetsya, chto attribute terminalnyx simvolov - libo
l
network
lexical analyzer.
symbol X as X (a). In the first case, the designation a (X) will be given later
test source
r
seti)
t
grammar dannoe utverjdenie budem schitat neaktualnÿm, t.e. for
is 'svod.dbf'). Analogichno vo vtorom sluchae: X (a) = 'svod.dbf'.
Set of personal attributes of important attributes:
obuchennyx setey)
rules for deducting attributes.
project
neobxodimy for some operational technologies:
sovokupnost dannyx)
ispolzuetsya dlya obucheniya or
testovaya vyborka
simvolov ne sushchestvuet semanticheskix pravil vyvoda dlya vÿchisleniya
atributnyx grammar, soglasno kotoromu ispolzovanie «chistogo» atributnogo formalizma
vÿzÿvaet trudnosti pri sozdanii translyatora.
sozdanie seti
source
predopredelennÿe constants, libo dostupny as rezultat raboty
name file set (obuchennoy
attribute grammar: 1) attribute a symbol X as a (X) or 2) attribute a
learning source
read network /
general source
In dannom sluchae v svyazi so slojnostyu i neodnoznachnostyu traktovki
symbol «=» (example, a (X) = 'svod.dbf' - attribute attribute to symbol X
read project /
create network
the name of the project (contains the series
istochnik dannyx imya file dannyx, kotoryy
dannoy grammatiki i dlya terminalnyx simvolov budut sushchestvovat
With uchetom dannyx rassuzhdeniy postroim nabor atributov, kotorye
obshchaya (generalnaya
set
Dannoe predpolojenie analogichno opytu razrabotki translyatorov dlya
Obÿchno v atributnyx grammatikax schitaetsya, chto dlya terminalnyx
obuchayushchaya vÿborka
chtenie (seti, proekta) or
testing (in DBF or DB format)
proekt
86
Machine Translated by Google

---

(8)
operating system
S ÿ BMAK | BMK
(4)
(10)
report
A ÿ AMA | AMAK | AM
L ÿ lm L | lp L | lr L | ln L | ls L | lk L | lc L | lb L
(12)
A ÿ i A | CA | v A | s A | TA
ispolzuem sleduyushchiy vid zapisi: sleva - traditsionnÿe pravila vyvoda
(3)
(9)
neterminalnye simvoly v kotoryx imeyut atributy), ot tsentra sprava - semantic rules for
attributes, po pravomu krayu - nomer pravila (v
M | BMA | BA | ÿ
(5)
T ÿ tt | tv | tu | tp
C | b2 C | b3 C | b4 C
on the results of training,
testing, vspomogatelnyx
Postroim predvaritelnye pravila s using attributes:
(7)
l (li) = 'network'
report
(11)
(1)
C ÿ ci | cs | cn | cu | ce | b1 | b2 | b3 |
l (s) = 'project'
l (s) = 'network'
l (i) = 'report' l
(v) = 'report'
C ÿ ÿ
Pri opisanii pravil vyvoda i semanticheskix pravil dlya atributov
(2)
l (L) = l (li)
l (rd) = 'source'
t (rd) = 'source'
L ÿ lm | lp | lr | ln | ls | lk | lc | lb
slov yazÿka (ukazÿvayutsya v osnovnom te pravila, terminalnÿe ili
K ÿ BMK | BMAK | BAK | B
A ÿ i | C | v | s | T
sootvetstvii s numeratsiey allowed pervonachalnoy grammar).
svodnaya otchetnaya information
(6)
B ÿ rd P | I rd P
b4
C ÿ ci C | cs C | cn C | cu C | ce C | b1
87
Machine Translated by Google

---

(15)
AI (v) = {l}
AI (rp) = {r}
mnojestvo nasleduemyx atributov):
As (L) = {l}
multiplicity of synthesized attributes; analogichno AI (li) = {l} - attribute l simvola li vxodit vo
mnojestvo nasleduemyx simvolov; AI (L) = Ø - the number of inherited attributes for the
symbol L is empty.
Zapis As (L) = {l} oboznachaet, chto attribute l simvola L vxodit vo
(14)
For dannoy grammar (As - mnojestvo sinteziruemyx atributov, Ai -
P ÿ mt f EN | mt f N | EN | N
(17)
AI (s) = {l}
(16)
g (rd) = 'source'
As (s) = Ø
AI (li) = {l}
AI (L) = Ø
AI (i) = {l}
AI (rd) = {l, t, g}
I ÿ d D | d
r (mk) = 'network'
r (rp) = 'project' r
(rp) = 'network'
I ÿ ÿ
As (i) = Ø
AI (mk) = {r}
(19)
(13)
As (v) = Ø
(18)
As (li) = Ø
N ÿ nl | na | nd | nn | ns | ns nl
As (rp) = Ø
D ÿ st sy | st
E ÿ ez | ec | ea | em | ex
M ÿ mk L | rp L | rp
As (mk) = Ø
As (rd) = Ø
88
Machine Translated by Google

---

lkor
vlojennosti) urovnem detalizatsii terminalnogo alfavita. Here
Constanta
sozdannoy seti
(p), kotoryy budet soderjat v sebe eto podmnojestvo. Vozmojno, predposledniy parameter
sleduet peredat ne tolko na etap obucheniya, no i
«Nespetsialista»), but dopolnitelno formiruetsya eshche ryad atributov.
const
func
obuchennosti seti
menyat svoe znachenie pri testirovanii.
vid functions
dannogo urovnya detalizatsii:
layers
obuchilas right on all
Poslednee znachenie attribute (lkor) otnositsya, skoree, k operatsii
p
chislo sloev
uroven tochnosti chislovoe znachenie,
ispolzuyushcheesya pri obuchenii
testirovanii po obuchayushchey
number of neurons in sloe
Dannÿy uroven harakizuetsya bolee glubokim (po stepeni
Significance of attributes, characteristics for this level of detail:
otklonenie dlya polya otveta
Dannÿy nabor znacheniy attribute (za isklyucheniem poslednego) otnositsya k operatsii
sozdaniya seti (mk), dlya etogo vvedem «obshchiy» attribute
soxranyayutsya vse atributy i pravila vyvoda predlojennÿe ranee (na urovne
at the stage of testing, although the technological constant is not necessary
numerical parameter
neuron
priznak
vid functions
I present some additional attributes, characteristics for
activation
pokazatel togo, chto set
to the end
primeram vyborki (100%
pravilnosti primerov pri
network params
chislo sloev sozdannoy seti
ver
obucheniya (L) i doljen otnositsya k mnojestvu sinteziruemyx atributov
set parameters (number of layers,
neurons, etc.)
2.2.2.2. Uroven "specialist" (special uroven)
number of neurons
seti, as dopustimoe
vyborke)
89
Machine Translated by Google

---

As (mk) = Ø
(As). Dannÿy priznak formiruetsya pri okonchanii tsikla obucheniya (1 variant
etapy nujno nachinat snachala) (b (A) = ÿ).
Correct rule (selection):
(17)
Dannyy podkhod (ispolzovanie atributivnoy grammatiki) rasshiritelno mojno
primenyat dlya razrabotki spetsializirovannyx
For dannoy grammar:
setey (at the request of the user), vÿpolnyat other operatsii, and then
M ÿ mk L | rp L | rp
(8)
- sravnivaetsya obshchee kolichestvo primerov i primerov, po kotorÿm set
(9)
AI (mk) = {p}
L ÿ lm L | lp L | lr L | ln L | ls L | lk L | lc L | lb
L
neuroimitatoru, kotoryye budut obuchat trebuemoe kolichestvo odnotipnyx
deystviy. Ispolzovanie gotovyx neuroimitatorov trebuet znachitelnyx
vÿdavat polzovatelyu svodnÿe rezultaty i opisaniya vÿpolnennyx
A ÿ i A | CA | v A | s A | TA
A ÿ i | C | v | s | T
obucheniya) i v sluchae lojnosti slovo yazÿka obrÿvaetsya (tehnologicheskie
AI (M) = {p}
obuchilas; 2 variant - opredelyayutsya percentage correctly opredelennyx
primerov na etape testirovaniya po vyborke, kotoraya ispolzovalas dlya
L ÿ lm | lp | lr | ln | ls | lk | lc | lb
As (M) = {l}
vremennyx zatrat dlya provedeniya issledovaniya, tak kak polnotsennÿe
vspomogatelnyx programs (modules, macros), podklyuchaemyx k
l (M) = p (mk) l (L)
p (mk) = 'const' p
(mk) = 'layers' p
(mk) = 'neuron' p
(mk) = 'func' p (mk)
= 'ver'
(7)
2.2.2.3. Uroven "developer" (uroven realizatsii)
(10)
90
Machine Translated by Google

---

(operatsii normirovaniya vÿpolnyayutsya vneshnimi programmami). Iz etix
kajdoy setyu, t. k. v neuroimitatorax ne predusmotrena vozmojnost
obrazom, mojno odnovremenno otkrÿt obuchayushchuyu i testovuyu vÿborki, a
form: 1) menu; 2) buttons in the toolbar ("hot" keys); 3) elements of settings. In
elementax settings can be
ustranit this osobennost, sushchestvenno avtomatizirovat i oblegchit
razrabotke dannogo yazyka i grammatiki ispolzovalsya drugoy podkhod, kotoryy
predpolagaet otkrÿtie tolko odnogo fayla dannyx (libo
struktury seti, opredelenie tipov poley (vxodnÿe, vyxodnÿe), znachenie
predobrabotki pri postroenii slov yazyka.
testirovanii), setting algorithms for training and methods of normalization. Elements
of the menu predstavlyayut operatsii, kotoryye ispolzuyutsya medium
Prisutstvie v alfavite otdelnyx terminalnyx simvolov dlya chteniya
These features opredelyayutsya before all interfaces
operatsiy, realizovannyx v neyroimitatore. In some zarubejnyx
rekomenduemÿy perechen operatsiy i ix raspolojenie na forme, v
operatsii, kotorye ispolzuyutsya ochen chasto: obuchenie seti, testirovanie,
opredelenie znachimostey. Operatsii normirovaniya vÿpolnyayutsya in nekoryx
neyroimitatorax
uslovno operatsii mojno raspolojit na sleduyushchix elementax ekrannoy
setey (obÿchno prixoditsya mnogokratno povtoryat odni i te je operatsii nad
testovoy vyborok) predstavleny otdelno v vide dvux knopok. Takim
trebuyut podavat normirovannÿe znacheniya vo vxodnom fayle dannyx
paketnogo ispolzovaniya kajdoy operatsii). Development of macros will allow
then provodit dalneyshie operatsii bez povtornogo ix otkrÿtiya. Pri
raspolojit operatsii naimenee ispolzuemÿe: naprimer, zadanie
soobrajeniy dannÿe operatsii vklyuchenÿ v obyazatelnÿy perechen operatsiy
otkloneniya vyxodnogo parameter (ispolzuetsya pri obuchenii i
the process of conducting multifaceted experiments with multiple sets.
obuchayushchey, libo testovoy vÿborki).
raznyx vÿborok predpolagaet nalichie sootvetstvuyushchix otdelnyx
neuroimitator, poetomu dlya sistematizatsii opisaniya privedem
frequency: naprimer, operatsii uproshcheniya set (kontrastirovanie, binarization),
verbalization. In the video button on the panel can be organized
neyroimitatorax dannÿe operatsii (v chastnosti chteniya obuchayushchey i
ÿÿÿÿÿÿÿÿÿÿÿÿ trebuyut vÿpolneniya odnotipnyx operatsiy nad mnojestvom
zavisimosti ot stepeni vajnosti i chastotÿ ispolzovaniya. Otmetim, chto
avtomaticheski, v inyx po zaprosu polzovatelya, nekotorye neyroimitatory
91
Machine Translated by Google

---

milk metodologii IDEF3, realizuyushchiy scenario protsessov posredstvom
dalneyshee razvitie sleduet opredelit klass atributnyx grammatik i
9. Quality of training sets
1. Technology
Method s ispolzovaniem atributnyx grammatical analogichen po svoey