Federal gosudarstvennoe uchebnoe predpriyatie Chair of the System of Artificial Intelligence

Podacha input signals and snyatie output signals sets
k naxojdeniyu skalyarnogo proizvedeniya vektora vkhodov na sootvetstvuyushchie
trudoemkost vÿchisleniy po vsey seti ravna NH, gde N - chislo neyronov.
Dlya ukazannoy set trudoemkost vÿchisleniy ravna 5H.
nekotoromu diapazonu znacheniy. Naibolee chasto ispolzuemÿe nelineynye
Budem schitat, chto trudoemkost vÿchisleniy (zatrachivaemoe vremya, zagruzka
protsessora i t.d.) for kajdogo neyrona odinakova i ravna H. Togda
5 5
4
5 1
7 1
3
8 2
5
3
6 2
5
4 4
3 3
2
2 2
4
9 3
10 4
37
8 2 3 1
2 2
10 4
1 1
4 2
5 ÿw9 f 3 ÿ w5 f1 ÿ ÿ 2 ÿw3 x1 ÿ w4 x2 ÿ ÿ
7 1
ÿ
y ÿ ft
t ÿ wx ÿ wx 1 1 1 2 2
ÿ ÿ
ÿ ÿ
ÿ ÿ
y ÿ ft
ÿ ÿ wfwfwx ÿ wx ÿ wfwx ÿ wx
t ÿ wy ÿ wy
ÿ ÿ
ÿ ÿ
z ÿ y ÿ ft
z ÿ f w1 x1 ÿ w2 x2 ÿ w6 f ÿ
t ÿ wy ÿ wy
t ÿ wy ÿ wy
t ÿ wx ÿ wx 3 1
1
2
1
1
4 2
Machine Translated by Google

---

correctly obrabatÿvatsya interpretatorom answer. For example, pri
signal in ix istinnom diapazone velichin i poluchat ot seti vyxodnÿe
where max xi i min xi - sootvetstvenno maksimalnoe i minimalnoe
normirovanie ne menyayut, no esli predpolagaetsya, chto v dalneyshem
normirovat, naprimer, v diapazon znacheniy [-1,1] ili [0,1] v tex sluchayax, kogda
sushchestvenna polojitelnost, libo delat tak, chtoby vxodnye
answers.
users by ego otsenkam. Inymi slovami, eti velichiny doljny
vyborki.
Vyhodnÿe signal seti doljny normirovatsya v diapazon istinnyx
kajdaya i-ya component vxodnogo vector dannyx xi zamenyaetsya novoy
tak, chtoby razbros dannym byl edinichnym.
,
doljno sootvetstvenno menyat i pravilo normirovaniya dannyx. Obychno v
For setey-klassifikatorov normirovka vyxodnyx signalov ne nujna, poskolku
polzovatel poluchaet ne sobstvenno vyxodnÿe signalaly seti, a
sushchestvuyushchego na nastoyashchiy moment zadachnika i v dalneyshem
range. This privodit k tomu, chto obÿchno nelzya podavat seti vxodnye
,
class. Zdes kajdyy vyxodnoy vector zadachnika preobrazuetsya tak, mine
signals in the required range.
Poetomu pered podachey seti vxodnyx signalov ix neobxodimo
znachenie dlya dannoy komponenty, vÿchislennÿe po vsey obuchayushchey vÿborke.
According to these formulas pereschityvayutsya and components of vectors
postupyat silno otlichayushchiesya dannÿe, to znacheniya min i max zadayutsya
klassifikatsii na tri klassa i interpreter "pobeditel zabiraet vse"
vvoditsya v moment sozdaniya seti i v dalneyshem ne zaviset ot obuchayushchey
signaly ne slishkom silno vyhodili za predelÿ etix otrezkov. Normirovanie mojno vÿpolnit
po ukazannym nije formulam, v kotoryx
Mojno normirovat i po-drugomu, naprimer, pereschityvat vyborku
velichinoy. Normirovanie na diapazon [0,1]:
Here imeetsya odna slojnost. Lyuboe izmenenie obuchayushchey vÿborki
znacheniy po obrashchennÿm formulam.
normalization of the range [-1,1]:
funktsii neyronov takje obÿchno prinimayut znacheniya iz nekotorogo
kachestve minimalnyx i maksimalnyx znacheniy berut dannÿe iz
rezultat otneseniya situatsii interpreterrom otveta k tomu ili inomu
38
Machine Translated by Google

---

squares of the elements of the vector YY ' ( otsenka MNK), libo bolee
- -1.
nekotorogo funktsionala nevyazki mejdu vyxodnymi signalami seti i
The set of parameters vector sootvetstvuyushchix chastnyx proizvodnyx budet
spaces. Vmesto takix otsutstvuyushchix component dannyx mojno podavat
Minimize funktsii otsenki vÿpolnyaetsya s privlecheniem
(otsenki v dalneyshem) proizvoditsya by means of such podstroyki obuchaemyx
funktsii otsenki, v xode kotoroy neyronnaya set "obuchaetsya" davat
signalov X vÿdavala takoy otvet Y ', kotoryy bÿl bÿ po vozmojnosti blije
nekotorÿmy pravdopodobnymi znacheniyami. Poslednee yavlyaetsya
strukturu seti, podavaya na vyxod seti chastnÿe proizvodnÿe funktsii otsenki
by summarnomu gradientu (gradientu po polnomu zadachniku), chto uskoryaet
primerov N (t.e. dlya N par {Xi, Yi}, i = 1, .., N) dostigalsya minimum summarnoy
funktsionirovanie, my mojem poluchit dlya kajdogo podstroechnogo
Obuchenie neyronnoy seti na "zadachnike" oznachaet minimizatsiyu
vector, sootvetstvuyushchaya nomeru klassa, imeet znachenie 1, a ostalnÿe dve
In kachestve funktsii normy vÿstupaet pokomponentnaya summa
otsenki otsenki functions by defining this parameter or signal. For vectors
Drugoy problemoy yavlyaetsya situation, kogda v tablitse dannyx imeyutsya
gradientom funktsii otsenki, poetomu vozmojna gradientnaya optimization
signalami, kotoryye trebuetsya poluchit. Minimization functional nevyazki
spetsializirovannaya.
parameters ai neuronov seti, chtoby set na nekotory vxodnoy vector
nul, mojno isklyuchat nekomplektnÿe vektora iz obuchayushchey vÿborki, mojno pered
obucheniem seti reshat zadachu zapolneniya probelov v dannyx
gradientnyx methods of optimization. Preobrazovav po nekotorym pravilam
trebuemÿe otvety na podavaemÿe ey vxodnÿe signalaly. Set luchshe obuchat
po vyxodnÿm signalam i ispolzuya tak nazÿvaemoe dvoystvennoe
samostoyatelnoy, often vstrechayushcheysya and dostatochno slojnoy problemoy.
k trebuemomu otvetu Y. Inymi slovami, chtoby dlya zadachnika s chislom
obucheniya process, i primenyat spetsializirovannÿe algorithms
Obuchenie neyronnyx setey
class number will be coded vector from trex number, where component
otsenki functions.
parameter i dlya kajdogo vxodnogo signalala seti chastnÿe proizvodnÿe