{"id":18618,"date":"2022-06-21T13:06:30","date_gmt":"2022-06-21T11:06:30","guid":{"rendered":"https:\/\/kinit.sk\/transparent-models-vs-black-box-models\/"},"modified":"2023-10-12T14:29:12","modified_gmt":"2023-10-12T12:29:12","slug":"transparentne-modely-vs-modely-typu-cierna-skrinka","status":"publish","type":"post","link":"https:\/\/kinit.sk\/sk\/transparentne-modely-vs-modely-typu-cierna-skrinka\/","title":{"rendered":"Transparentn\u00e9 modely vs. modely typu \u010dierna skrinka"},"content":{"rendered":"
\n

Vysvetlite\u013en\u00e1 umel\u00e1 inteligencia: od \u010diernych skriniek k transparentn\u00fdm modelom<\/h2>\n<\/div>\n\n
\n

M\u00f4\u017ee umel\u00e1 inteligencia rozhodova\u0165 o tom, kto bude prepusten\u00fd z v\u00e4zenia, alebo kto post\u00fapi do \u010fal\u0161ieho kola pracovn\u00fdch pohovorov? M\u00f4\u017ee odpor\u00fa\u010da\u0165 prevent\u00edvne vy\u0161etrenia na r\u00f4zne choroby alebo \u00faplne prevzia\u0165 zodpovednos\u0165 za riadenie vozidla v snehovej b\u00farke? <\/p>\n<\/div>\n\n

\n

Nie ste si ist\u00ed, ako odpoveda\u0165 na tak\u00e9to ot\u00e1zky? Ak by ste vedeli, \u017ee umel\u00e1 inteligencia svoje rozhodnutie detailne zd\u00f4vodnila a ka\u017ed\u00fd \u010dlovek by si mohol overi\u0165, \u010di sa rozhodovala spr\u00e1vne \u2013 zmenilo by to v\u00e1\u0161 n\u00e1zor? <\/p>\n<\/div>\n\n

\n

Prin\u00e1\u0161ame s\u00e9riu \u010dl\u00e1nkov<\/strong> na t\u00e9mu vysvetlite\u013enej a transparentnej umelej inteligencie. <\/p>\n<\/div>\n\n

\n

V tretej \u010dasti sa pozrieme na dve kateg\u00f3rie modelov umelej inteligencie – transparentn\u00e9 modely a modely typu \u010dierna skrinka<\/strong>. Ak\u00e9 s\u00fa medzi nimi rozdiely a pre\u010do pou\u017e\u00edvame aj netransparentn\u00e9 modely, ke\u010f je schopnos\u0165 pochopi\u0165 ich predpovede tak\u00e1 d\u00f4le\u017eit\u00e1? Ktor\u00e9 modely m\u00f4\u017eeme pova\u017eova\u0165 za transparentn\u00e9 a ktor\u00e9 si, naopak, vy\u017eaduj\u00fa dodato\u010dn\u00fa interpret\u00e1ciu a vysvetlenie ich predpoved\u00ed?<\/p>\n<\/div>\n\n

\n

Transparentn\u00e9 modely<\/h3>\n<\/div>\n\n
\n

Transparentn\u00e9 modely umelej inteligencie s\u00fa tak\u00e9, pri ktor\u00fdch vieme pochopi\u0165 nielen to, ako funguj\u00fa po algoritmickej str\u00e1nke, ale aj to, pre\u010do sa rozhoduj\u00fa tak, ako sa rozhoduj\u00fa. Vieme teda ako \u013eudia pochopi\u0165, ako modely \u201cprem\u00fd\u0161\u013eaj\u00fa\u201d a ak\u00e9 znalosti sa z\u00edskali o re\u00e1lnom svete.<\/strong> Tieto modely s\u00fa typicky star\u0161ie, menej v\u00fdkonn\u00e9 a jednoduch\u0161ie*. <\/p>\n<\/div>\n\n

\n

Pr\u00edkladom transparentn\u00e9ho modelu s\u00fa rozhodovacie stromy. V rozhodovac\u00edch stromoch s\u00fa znalosti, na z\u00e1klade ktor\u00fdch model vytvor\u00ed predpove\u010f pre vstup, vyjadren\u00e9 explicitne ako pravidl\u00e1 (napr. \u201cAK m\u00e1 \u010dlovek pr\u00edjem viac ako 1000\u20ac A m\u00e1 menej ako 40 rokov, TAK je mo\u017en\u00e9 schv\u00e1li\u0165 mu \u00faver vo v\u00fd\u0161ke 30.000\u20ac\u201d).<\/p>\n<\/div>\n\n

\n

* Zauj\u00edmavos\u0165ou je, \u017ee aj tak\u00e9to prirodzene transparentn\u00e9 modely sa m\u00f4\u017eu pomerne jednoducho zmeni\u0165 na \u0165a\u017eko pochopite\u013en\u00e9. Sta\u010d\u00ed, ak je reprezent\u00e1cia d\u00e1t, ktor\u00e9 s\u00fa na ich vstupe dostato\u010dne komplexn\u00e1 alebo ak s\u00fa modely ve\u013emi ve\u013ek\u00e9. V takomto pr\u00edpade je pre \u010dloveka \u0165a\u017ek\u00e9 a\u017e nemo\u017en\u00e9 simulova\u0165 rozhodovac\u00ed proces modelu, \u010do je jedna z podmienok transparentnosti.<\/p>\n<\/div>\n\n

\n
\"\"
Obr\u00e1zok<\/em>: <\/strong>Zjednodu\u0161en\u00fd pr\u00edklad rozhodovacieho stromu, ktor\u00fd poskytuje odpor\u00fa\u010danie, \u010di by mala alebo nemala by\u0165 \u017eiadate\u013eovi poskytnut\u00e1 hypot\u00e9ka. Strom sa rozhoduje na z\u00e1klade \u0161tyroch atrib\u00fatov – veku a \u010dist\u00e9ho pr\u00edjmu \u017eiadate\u013ea, toho, \u010di m\u00e1 spoluru\u010dite\u013ea a pr\u00edpadne \u010dist\u00e9ho pr\u00edjmu spoluru\u010dite\u013ea. V tomto pr\u00edpade by strom odporu\u010dil \u017eiados\u0165 zamietnu\u0165. \u017diadate\u013eka m\u00e1 viac ako 40 rokov, jej pr\u00edjem je ni\u017e\u0161\u00ed ako 1300\u20ac a z\u00e1rove\u0148 spolo\u010dn\u00fd pr\u00edjem so spoluru\u010dite\u013eom je men\u0161\u00ed ako 1600\u20ac. Strom si vytv\u00e1ra pravidl\u00e1 po\u010das procesu u\u010denia. V tomto pr\u00edpade by sa napr\u00edklad mohol strom u\u010di\u0165 na z\u00e1klade toho, ktor\u00e9 skupiny \u017eiadate\u013eov v minulosti riadne spl\u00e1cali \u00favery a ktor\u00e9 nie. Tak\u00fdto model m\u00f4\u017eeme pova\u017eova\u0165 za transparentn\u00fd, preto\u017ee \u010dlovek dok\u00e1\u017ee jednozna\u010dne simulova\u0165 a pochopi\u0165, pre\u010do bolo rozhodnutie modelu tak\u00e9, ak\u00e9 bolo.<\/em><\/figcaption><\/figure>\n<\/div>\n\n
\n

Modely typu \u010dierna skrinka<\/h3>\n<\/div>\n\n
\n

Aj pre komplexn\u00e9 modely typu \u010dierna skrinka plat\u00ed, \u017ee vieme pochopi\u0165, ako pracuj\u00fa. Rozumieme ich mechanike a matematick\u00fdm oper\u00e1ci\u00e1m, ktor\u00e9 s\u00fa v procese u\u010denia vyu\u017e\u00edvan\u00e9. \u010comu teda nerozumieme?<\/p>\n<\/div>\n\n

\n

Modern\u00e9 modely typu \u010dierna skrinka s\u00fa nato\u013eko komplexn\u00e9, \u017ee nie je mo\u017en\u00e9 \u00faplne simulova\u0165 to, ako pri\u0161li k rozhodnutiu<\/strong>. Nie tak, aby bola t\u00e1to simul\u00e1cia pre \u010dloveka zrozumite\u013en\u00e1 nato\u013eko, aby pochopil, ak\u00e9 podnety a znalosti zak\u00f3dovan\u00e9 v modeli viedli k ich rozhodnutiu. In\u00fdmi slovami, nie je v \u013eudsk\u00fdch sil\u00e1ch vzia\u0165 si pero a papier a nasimulova\u0165 cel\u00fd rozhodovac\u00ed proces modelu, preto\u017ee je jednoducho pr\u00edli\u0161 komplexn\u00fd<\/strong>. Jednotliv\u00e9 oper\u00e1cie, ktor\u00e9 prebiehaj\u00fa v modeli, pritom m\u00f4\u017eu by\u0165 ve\u013emi jednoduch\u00e9. M\u00f4\u017ee sa jedna\u0165 napr\u00edklad o s\u010d\u00edtanie a n\u00e1sobenie. Ke\u010f v\u0161ak model obsahuje tak\u00fdchto oper\u00e1ci\u00ed mili\u00f3ny a\u017e miliardy, nie je v \u013eudsk\u00fdch sil\u00e1ch simulova\u0165 alebo aspo\u0148 skontrolova\u0165 efekt ka\u017edej z nich.<\/p>\n<\/div>\n\n

\n

Typick\u00fdm predstavite\u013eom tejto kateg\u00f3rie modelov s\u00fa hlbok\u00e9 neur\u00f3nov\u00e9 siete<\/strong>. Zatia\u013e \u010do v rozhodovac\u00edch stromoch s\u00fa znalosti vyjadren\u00e9 explicitne ako pravidl\u00e1, v pr\u00edpade neur\u00f3nov\u00fdch siet\u00ed s\u00fa znalosti zap\u00edsan\u00e9 vo forme zv\u00e4\u010d\u0161a mali\u010dk\u00fdch \u010d\u00edseln\u00fdch hodn\u00f4t – parametrov modelu<\/em>. Predpove\u010f je potom vytvoren\u00e1 v\u010faka postupnosti matematick\u00fdch oper\u00e1ci\u00ed (naj\u010dastej\u0161ie nad maticami a vektormi), ktor\u00e9 transformuj\u00fa vstupn\u00e9 d\u00e1ta na fin\u00e1lnu predpove\u010f modelu. <\/p>\n<\/div>\n\n

\n

Ilustr\u00e1ciu toho, ako neur\u00f3ny v skryt\u00fdch vrstv\u00e1ch kombinuj\u00fa a postupne filtruj\u00fa vstupy za pomoci parametrov, m\u00f4\u017eete vidie\u0165 na obr\u00e1zku z prvej \u010dasti s\u00e9rie<\/a>. Do obr\u00e1zku sme doplnili pr\u00edklad v\u00fdpo\u010dtu v\u00fdstupu neur\u00f3nu na prvej skrytej vrstve. Je vidno, \u017ee v\u00fdpo\u010det hodnoty jedn\u00e9ho neur\u00f3nu naozaj nie je \u017eiadna raketov\u00e1 veda a sta\u010d\u00ed na to vedie\u0165 spo\u010d\u00edtava\u0165 a n\u00e1sobi\u0165.<\/p>\n<\/div>\n\n

\n
\"\"
Obr\u00e1zok<\/em>: <\/strong>Pr\u00edklad v\u00fdpo\u010dtu v\u00fdstupu jedn\u00e9ho neur\u00f3nu v r\u00e1mci neur\u00f3novej siete. V r\u00e1m\u010deku je mo\u017en\u00e9 vidie\u0165, ako m\u00f4\u017ee vyzera\u0165 v\u00fdpo\u010det hodnoty jedn\u00e9ho zo skryt\u00fdch neur\u00f3nov na prvej skrytej vrstve. \u201ca\u201d, \u201cb\u201d, \u201cc\u201d a \u201cBIAS\u201d s\u00fa pr\u00edklady parametrov neur\u00f3novej siete – pr\u00e1ve ony tvoria jej pam\u00e4\u0165 a t\u00fdm, \u017ee neur\u00f3nov\u00e1 sie\u0165 nastavuje ich hodnoty v procese u\u010denia, sa u\u010d\u00ed. Dne\u0161n\u00e9 re\u00e1lne pou\u017e\u00edvan\u00e9 neur\u00f3nov\u00e9 siete m\u00f4\u017eu ma\u0165 stovky tis\u00edc a\u017e stovky mili\u00e1rd parametrov.<\/em><\/figcaption><\/figure>\n<\/div>\n\n
\n

Ako \u010fal\u0161\u00ed pr\u00edklad sa pozrieme na klasifik\u00e1ciu \u010diernobielych obr\u00e1zkov do piatich r\u00f4znych tried – auto, n\u00e1kladiak, lietadlo, lo\u010f a k\u00f4\u0148. Obr\u00e1zok m\u00f4\u017eeme v po\u010d\u00edta\u010di reprezentova\u0165 ako dvojdimenzion\u00e1lnu mrie\u017eku \u010d\u00edsel – maticu, v ktorej ka\u017ed\u00e1 bunka obsahuje mal\u00e9 \u010d\u00edslo. Toto \u010d\u00edslo reprezentuje hodnotu konkr\u00e9tneho pixelu (pri \u010diernobielom obr\u00e1zku je biela farba reprezentovan\u00e1 naj\u010dastej\u0161ie \u010d\u00edslom 255, zatia\u013e \u010do \u010dierny pixel m\u00e1 hodnotu 0). <\/p>\n<\/div>\n\n

\n

Neur\u00f3nov\u00e1 sie\u0165 v prvej vrstve za pomoci r\u00f4znych jednoduch\u00fdch matematick\u00fdch oper\u00e1ci\u00ed, ako s\u00fa n\u00e1sobenie a s\u010d\u00edtanie, skombinuje tieto \u010d\u00edsla so svojimi vlastn\u00fdmi parametrami. V\u010faka tomu sa tento obr\u00e1zok zmen\u00ed na in\u00fa mrie\u017eku \u010d\u00edsel. Tieto \u010d\u00edsla u\u017e ale nie s\u00fa hodnotami pixelov samotn\u00e9ho obr\u00e1zku, ale s\u00fa ak\u00fdmsi filtr\u00e1tom (odborne hovor\u00edme o skrytej reprezent\u00e1cii). Pri pr\u00e1ci s obrazov\u00fdmi d\u00e1tami sa na to v neur\u00f3nov\u00fdch sie\u0165ach \u010dasto vyu\u017e\u00edva oper\u00e1cia konvol\u00facia, ktor\u00e1 je napr\u00edklad z\u00e1kladom mnoh\u00fdch filtrov v obr\u00e1zkov\u00fdch editoroch (napr. Photoshop).<\/p>\n<\/div>\n\n

\n
\"\"
Obr\u00e1zok<\/em>: <\/strong>V neur\u00f3nov\u00fdch sie\u0165ach sa pri pr\u00e1ci s obrazov\u00fdmi d\u00e1tami \u010dasto pou\u017e\u00edva oper\u00e1cia konvol\u00facia (odtia\u013e n\u00e1zov konvolu\u010dn\u00e9 neur\u00f3nov\u00e9 siete). Na rovnakom princ\u00edpe pracuj\u00fa mnoh\u00e9 popul\u00e1rne filtre v editoroch fotografi\u00ed. Ak ste v takomto editore niekedy napr\u00edklad vyhladili \u010di rozmazali fotografiu, s ve\u013ekou pravdepodobnos\u0165ou ste pou\u017eili filter zalo\u017een\u00fd na konvol\u00facii.<\/em><\/figcaption><\/figure>\n<\/div>\n\n
\n

Druh\u00e1 skryt\u00e1 vrstva n\u00e1sledne urob\u00ed presne to ist\u00e9, \u010do prv\u00e1, no na jej vstupe u\u017e nie s\u00fa pixely obr\u00e1zka, ale mrie\u017eka, ktor\u00e1 vznikla v prvom kroku. Tento proces sa opakuje, a\u017e k\u00fdm v neur\u00f3novej sieti ned\u00f4jdeme k poslednej, v\u00fdstupnej vrstve, ktorej v\u00fdstupom s\u00fa \u010d\u00edsla. Ka\u017ed\u00e9 z t\u00fdchto \u010d\u00edsiel n\u00e1m povie, do akej miery \u201cje model presved\u010den\u00fd\u201d, \u010di je na obr\u00e1zku auto, n\u00e1kladiak, lietadlo, lo\u010f alebo k\u00f4\u0148. Ak je najv\u00e4\u010d\u0161ie prv\u00e9 \u010d\u00edslo z tejto p\u00e4tice, model je presved\u010den\u00fd, \u017ee na obr\u00e1zku sa nach\u00e1dza auto, ak druh\u00e9, bude to n\u00e1kladiak at\u010f..<\/p>\n<\/div>\n\n

\n

Ako vid\u00edme, oproti rozhodovac\u00edm stromom je pri neur\u00f3nov\u00fdch sie\u0165ach ve\u013emi n\u00e1ro\u010dn\u00e9 poveda\u0165, pre\u010do bola predpove\u010f modelu tak\u00e1, ak\u00e1 bola. \u010casto iba nepriamo pozorujeme, ako sa vstup (napr. obr\u00e1zok) vo vn\u00fatri modelu transformuje z jednej skrytej reprezent\u00e1cie na in\u00fa a na z\u00e1klade toho m\u00f4\u017eeme odhadova\u0165, \u010do sa vo vn\u00fatri modelu deje. <\/p>\n<\/div>\n\n

\n

Pr\u00edklad postupnej filtr\u00e1cie od vstupn\u00e9ho obr\u00e1zku a\u017e po v\u00fdsledn\u00fa predpove\u010f siete m\u00f4\u017eete vidie\u0165 na obr. 3. Ako vidno, v\u00fdstupy prv\u00fdch vrstiev neur\u00f3novej siete (CONV, RELU, \u2026) s\u00fa pre \u010dloveka dobre \u010ditate\u013en\u00e9 a dokonca m\u00f4\u017eeme odhadn\u00fa\u0165, na \u010do sa konkr\u00e9tna vrstva zameriava. Napr\u00edklad na v\u00fdstupoch druhej vrstvy (RELU) vid\u00edme, \u017ee niektor\u00e9 \u010dasti siete sa zamerali na detekciu tie\u0148a pod autom, zatia\u013e \u010do in\u00e9 detegovali okolit\u00fd les. <\/p>\n<\/div>\n\n

\n

\u010c\u00edm hlb\u0161ie ale v sieti ideme, t\u00fdm men\u0161iu schopnos\u0165 interpretova\u0165, \u010do sa vo vn\u00fatri siete deje, m\u00e1me. Napr\u00edklad nevieme, \u010do znamenaj\u00fa svetl\u00e9 oblasti, ktor\u00e9 s\u00fa v\u00fdstupom z poslednej skrytej vrstvy siete (POOL). Znamen\u00e1 svetl\u00fd pixel, \u017ee sa na obr\u00e1zku nach\u00e1dza koleso? Alebo to znamen\u00e1, \u017ee je auto oto\u010den\u00e9 smerom doprava?<\/p>\n<\/div>\n\n

\n
\"\"
Obr\u00e1zok 3<\/em><\/a><\/figcaption><\/figure>\n<\/div>\n\n
\n

E\u0161te raz – \u010domu teda ako \u013eudia nerozumieme? Pr\u00edklad jazykov\u00e9ho modelu na anal\u00fdzu sentimetu<\/h3>\n<\/div>\n\n
\n

Pozrime sa na konkr\u00e9tny pr\u00edklad – umel\u00fa neur\u00f3nov\u00fa sie\u0165, ktorej \u00falohou je poveda\u0165, \u010di m\u00e1 text pozit\u00edvny alebo negat\u00edvny sentiment.<\/p>\n<\/div>\n\n

\n

Pre\u010do bolo konkr\u00e9tne rozhodnutie tak\u00e9, ak\u00e9 bolo?<\/h4>\n<\/div>\n\n
\n

Ak dostato\u010dne ve\u013ek\u00fd jazykov\u00fd model ako BERT*, \u010do je v skuto\u010dnosti mohutn\u00e1 umel\u00e1 neur\u00f3nov\u00e1 sie\u0165, natr\u00e9nujeme na textov\u00fdch d\u00e1tach s negat\u00edvnym a pozit\u00edvnym sentimentom, tento model bude fungova\u0165, a zrejme aj presne. <\/p>\n<\/div>\n\n

\n

Nevieme v\u0161ak, na z\u00e1klade \u010doho model prirad\u00ed textu pozit\u00edvny alebo negat\u00edvny sentiment. Pova\u017euje model text za negat\u00edvny kv\u00f4li tomu, \u017ee obsahuje prive\u013ea v\u00fdkri\u010dn\u00edkov? Alebo je to kv\u00f4li slovu \u201cpeklo\u201d, ktor\u00e9 autor textu pou\u017eil? <\/p>\n<\/div>\n\n

\n

Aj tak\u00e1to na prv\u00fd poh\u013ead jednoduch\u00e1 \u00faloha vy\u017eaduje, aby model ve\u013emi dobre rozumel kontextu a aj samotn\u00e9mu jazyku. Je rozdiel medzi recenziou \u201c\u010ca\u0161n\u00edci by mali \u00eds\u0165 do pekla!!!\u201d a \u201cV tejto re\u0161taur\u00e1cii pekelne dobre varia!!!\u201d<\/strong>.<\/p>\n<\/div>\n\n

\n

* BERT: angl. Bidirectional Encoder Representations from Transformers. V s\u00fa\u010dasnosti jeden z najpou\u017e\u00edvanej\u0161\u00edch a najsilnej\u0161\u00edch modelov na spracovanie prirodzen\u00e9ho jazyka.<\/p>\n<\/div>\n\n

\n

Ak\u00fa znalos\u0165 model obsahuje a kde je t\u00e1 znalos\u0165 ulo\u017een\u00e1?<\/h4>\n<\/div>\n\n
\n

Ak natr\u00e9novan\u00fd jazykov\u00fd model funguje dobre, je zrejm\u00e9, \u017ee do nejakej miery rozumie \u013eudsk\u00e9mu jazyku. Je v\u0161ak n\u00e1ro\u010dn\u00e9 poveda\u0165, \u010di a do akej miery rozumie jednotliv\u00fdm aspektom jazyka<\/strong>. Rozumie model rozdielu medzi slovesami a podstatn\u00fdmi menami? Nau\u010dil sa gramatiku a pravidl\u00e1 skladby viet v konkr\u00e9tnom jazyku? Ch\u00e1pe v\u00fdznam (s\u00e9mantiku) jednotliv\u00fdch slov, alebo sa len nau\u010dil \u0161tatistiku ich vz\u00e1jomn\u00e9ho v\u00fdskytu?<\/p>\n<\/div>\n\n

\n

Pri na\u0161ej modelovej \u00falohe, anal\u00fdze sentimentu, sa m\u00f4\u017eeme dovt\u00edpi\u0165, \u017ee sa model okrem in\u00e9ho koncentruje na pozit\u00edvne a negat\u00edvne slov\u00e1, ale nevieme, kde je t\u00e1to znalos\u0165 ulo\u017een\u00e1. Je to konkr\u00e9tny neur\u00f3n, ktor\u00fd je zodpovedn\u00fd za to, \u017ee v texte vieme rozozna\u0165 pozit\u00edvny sentiment? Alebo je to kombin\u00e1cia viacer\u00fdch neur\u00f3nov? Ako sme uk\u00e1zali v predch\u00e1dzaj\u00facej \u010dasti, interpretova\u0165 v\u00fdznam jednotliv\u00fdch parametrov modelu a ich vplyvu na v\u00fdsledn\u00fa predpove\u010f je netrivi\u00e1lny probl\u00e9m.<\/p>\n<\/div>\n\n

\n

S\u00fa transparentn\u00e9 modely m\u0155tve?<\/h2>\n<\/div>\n\n
\n

V\u00e4\u010d\u0161inu modelov, o ktor\u00fdch m\u00f4\u017eeme prehl\u00e1si\u0165, \u017ee s\u00fa transparentn\u00e9, tvoria jednoduch\u0161ie modely (rozhodovac\u00ed proces v komplexn\u00fdch modeloch alebo modeloch s prive\u013ea pravidlami \u010dlovek nedok\u00e1\u017ee kognit\u00edvne obsiahnu\u0165) s men\u0161ou modelovacou kapacitou. Mohlo by sa teda zda\u0165, \u017ee ich vyu\u017eitie je mal\u00e9 a na zlo\u017eitej\u0161ie \u00falohy nesta\u010dia. <\/p>\n<\/div>\n\n

\n

Pravdou v\u0161ak je, \u017ee transparentn\u00e9 modely sa st\u00e1le v praxi pou\u017e\u00edvaj\u00fa<\/strong>. A to z toho d\u00f4vodu, \u017ee s\u00fa oblasti, v ktor\u00fdch je transparentnos\u0165 minim\u00e1lne tak\u00e1 d\u00f4le\u017eit\u00e1, ako samotn\u00e1 v\u00fdkonnos\u0165 modelu<\/strong>. V prest\u00ed\u017enom \u010dasopise Nature ned\u00e1vno vy\u0161iel \u010dl\u00e1nok [1], ktor\u00fd kritizuje bezd\u00f4vodn\u00e9 nadpou\u017e\u00edvanie pr\u00edli\u0161 komplexn\u00fdch modelov, aj ke\u010f by na rie\u0161enie dan\u00e9ho probl\u00e9mu \u00faplne posta\u010dovali jednoduch\u0161ie, no zato transparentn\u00e9 modely.<\/p>\n<\/div>\n\n

\n

Pr\u00edklad: Nature Compas\/Corels<\/h4>\n<\/div>\n\n
\n

Compas je program, ktor\u00fd pou\u017e\u00edvaj\u00fa viacer\u00e9 s\u00fady v USA na odhadnutie pravdepodobnosti, \u017ee \u010dlovek, ktor\u00fd sp\u00e1chal trestn\u00fd \u010din, bude znovu kona\u0165 protiz\u00e1konne<\/strong>. Je to program, ktor\u00fd priamo ovplyv\u0148uje \u017eivoty \u013eud\u00ed, no z\u00e1rove\u0148 je vyslovene netransparentn\u00fd. Spolo\u010dnos\u0165, ktor\u00e1 program vyv\u00edja, dokonca ani nezverejnila princ\u00edp, na akom funguje.<\/p>\n<\/div>\n\n

\n

Vo vy\u0161\u0161ie spom\u00ednanom \u010dl\u00e1nku v \u010dasopise Nature autorka Cynthia Rudin uviedla alternat\u00edvnu met\u00f3du Corels, ktor\u00e1 m\u00e1 ove\u013ea menej parametrov, je transparentnej\u0161ia a pritom rovnako presn\u00e1 ako Compas.<\/p>\n<\/div>\n\n

\n

Je prirodzen\u00e9, \u017ee \u010dlovek, o ktorom sa rozhoduje, \u010di p\u00f4jde do v\u00e4zenia alebo nie, chce pozna\u0165 d\u00f4vod, pre ktor\u00fd sa s\u00fad rozhodol tak, ako sa rozhodol. Aj \u010dlovek, ktor\u00fd sp\u00e1chal nejak\u00fd priestupok, m\u00e1 toti\u017e pr\u00e1vo presved\u010di\u0165 sa, \u017ee proces a d\u00f4vody, pre ktor\u00e9 poputuje do v\u00e4zenia, s\u00fa f\u00e9r<\/strong>. <\/p>\n<\/div>\n\n

\n

Aby to nebolo pr\u00edli\u0161 jednoduch\u00e9, vieme sa na tento konkr\u00e9tny probl\u00e9m pozrie\u0165 aj z in\u00e9ho poh\u013eadu. Bolo by naozaj dobr\u00e9, ak by potenci\u00e1lni zlo\u010dinci presne vedeli, za ak\u00fdch okolnost\u00ed ich softv\u00e9r ozna\u010d\u00ed ako vysoko rizikov\u00fdch a za ak\u00fdch nie?<\/strong> Nemohlo by to vies\u0165 k snahe oklama\u0165 syst\u00e9m za \u00fa\u010delom dosiahnutia ni\u017e\u0161ieho trestu?<\/p>\n<\/div>\n\n

\n
\n

V dom\u00e9nach, kde sa rozhoduje o \u013eudsk\u00fdch osudoch, zaho\u010fme komplexn\u00e9 modely a pou\u017e\u00edvajme tie jednoduch\u0161ie, ale pochopite\u013enej\u0161ie, ak s nimi vieme dosiahnu\u0165 dostato\u010dn\u00fa presnos\u0165.<\/em><\/p>\n<\/div><\/blockquote>\n<\/div>\n\n

\n

Napriek nedostatkom, ktor\u00e9 vypl\u00fdvaj\u00fa z komplexnosti, st\u00e1le plat\u00ed, \u017ee najsilnej\u0161ie modely, ktor\u00e9 m\u00e1me, s\u00fa z\u00e1rove\u0148 najkomplexnej\u0161ie a \u0165a\u017eko vysvetlite\u013en\u00e9 a interpretovate\u013en\u00e9<\/strong>. V prvej \u010dasti seri\u00e1lu sme si uk\u00e1zali, \u017ee vo\u013eba algoritmu strojov\u00e9ho u\u010denia pre rie\u0161enie konkr\u00e9tneho probl\u00e9mu je \u010dasto o h\u013eadan\u00ed kompromisov. Zatia\u013e \u010do komplexn\u00e9 modely zvy\u010dajne dosahuj\u00fa vy\u0161\u0161iu presnos\u0165, miera ich interpretovate\u013enosti b\u00fdva ni\u017e\u0161ia. V nasledovnej \u010dasti si predstav\u00edme met\u00f3dy, ktor\u00e9 by n\u00e1m mohli pom\u00f4c\u0165 zv\u00fd\u0161i\u0165 \u00farove\u0148 ich transparentnosti.<\/p>\n<\/div>\n\n

\n

Z\u00e1ver<\/h2>\n<\/div>\n\n
\n

V tomto dieli seri\u00e1lu o vysvetlite\u013enej umelej inteligencii sme sa pozreli na dve kateg\u00f3rie modelov umelej inteligencie – transparentn\u00e9 modely a modely typu \u010dierna skrinka. Pri transparentn\u00fdch modeloch sme ako \u013eudia schopn\u00ed pochopi\u0165 nielen mechanizmus ich fungovania, ale aj proces ich rozhodovania<\/strong>. Vieme teda vysvetli\u0165, pre\u010do ich rozhodnutia boli tak\u00e9, ak\u00e9 boli. <\/p>\n<\/div>\n\n

\n

\u010cierne skrinky s\u00fa typicky komplexn\u00e9 modely, ktor\u00fdch rozhodnutia nie je mo\u017en\u00e9 priamo vysvetli\u0165 a v ktor\u00fdch s\u00fa znalosti ukladan\u00e9 vo forme, ktor\u00e1 nie je zrozumite\u013en\u00e1 pre \u010dloveka<\/strong>. Ak chceme zv\u00fd\u0161i\u0165 ich transparentnos\u0165, mus\u00edme na ne aplikova\u0165 dodato\u010dn\u00e9 mechanizmy vysvetlite\u013enosti a interpretovate\u013enosti.<\/p>\n<\/div>\n\n

\n


V \u010fal\u0161\u00edch dieloch sa pozrieme na konkr\u00e9tne pr\u00edklady met\u00f3d vysvetlite\u013enosti a interpretovate\u013enosti<\/strong>. Uk\u00e1\u017eeme si, pre\u010do n\u00e1m nesta\u010d\u00ed len jedna met\u00f3da vysvetlite\u013enosti.  Taktie\u017e sa dozviete, ako m\u00f4\u017eeme mera\u0165 kvalitu vysvetlenia predpovede umelej inteligencie a povieme si, ak\u00e9 benefity m\u00f4\u017eeme ma\u0165 ako \u013eudia z  pou\u017e\u00edvania transparentn\u00fdch a vysvetlite\u013en\u00fdch met\u00f3d umelej inteligencie.<\/p>\n<\/div>\n\n

\n
\n
\n

<\/p>\n<\/div>\n\n

\n

<\/p>\n<\/div>\n\n

\n

Projekt podporil Nada\u010dn\u00fd fond PricewaterhouseCoopers v Nad\u00e1cii Pontis.<\/em><\/p>\n<\/div><\/div>\n<\/div>\n\n

\n
\n
\"\"<\/figure>\n<\/div><\/div>\n<\/div><\/div>\n<\/div>\n\n
\n

Zdroje<\/strong>:<\/p>\n<\/div>\n\n

\n

[1] RUDIN, Cynthia. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence<\/em>, 2019, 1.5: 206-215.<\/p>\n<\/div>\n\n

\n\t
\n\t\t
\n\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t

Vysvetlite\u013en\u00e1 umel\u00e1 inteligencia: od \u010diernych skriniek k transparentn\u00fdm modelom<\/h2><\/div>\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\"\"\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t

Ako rozm\u00fd\u0161\u013ea umel\u00e1 inteligencia?<\/h3>\n \t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\tapr 28. 2022\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t
Vysvetlite\u013en\u00e1 umel\u00e1 inteligencia: od \u010diernych skriniek k transparentn\u00fdm modelom M\u00f4\u017ee umel\u00e1 inteligencia rozhodova\u0165 o tom, kto bude prepusten\u00fd z v\u00e4zenia, alebo kto post\u00fapi do \u010fal\u0161ieho kola pracovn\u00fdch pohovorov? M\u00f4\u017ee odpor\u00fa\u010da\u0165…<\/div>\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t\t\t<\/a>\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t<\/article>\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\"\"\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t

Tri pr\u00ednosy vysvetlite\u013enej umelej inteligencie<\/h3>\n \t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\tm\u00e1j 25. 2022\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t
Vysvetlite\u013en\u00e1 umel\u00e1 inteligencia: od \u010diernych skriniek k transparentn\u00fdm modelom Umel\u00e1 inteligencia bude \u010doraz v\u00e4\u010d\u0161mi zasahova\u0165 do na\u0161ich \u017eivotov. Ve\u013ea \u013eud\u00ed v\u0161ak umelej inteligencii ned\u00f4veruje a maj\u00fa probl\u00e9m prenecha\u0165 jej aj…<\/div>\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t\t\t<\/a>\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t<\/article>\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\"\"\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t

Transparentn\u00e9 modely vs. modely typu \u010dierna skrinka<\/h3>\n \t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\tj\u00fan 21. 2022\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t
Vysvetlite\u013en\u00e1 umel\u00e1 inteligencia: od \u010diernych skriniek k transparentn\u00fdm modelom M\u00f4\u017ee umel\u00e1 inteligencia rozhodova\u0165 o tom, kto bude prepusten\u00fd z v\u00e4zenia, alebo kto post\u00fapi do \u010fal\u0161ieho kola pracovn\u00fdch pohovorov? M\u00f4\u017ee odpor\u00fa\u010da\u0165…<\/div>\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t\t\t<\/a>\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t<\/article>\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\"\"\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t

Komponenty a vlastnosti dobr\u00fdch vysvetlen\u00ed rozhodnut\u00ed modelov umelej inteligencie<\/h3>\n \t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\taug 16. 2022\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t
Vysvetlite\u013en\u00e1 umel\u00e1 inteligencia: od \u010diernych skriniek k transparentn\u00fdm modelom Z poh\u013eadu strojov\u00e9ho u\u010denia (angl. Machine learning, ML) \u017eijeme \u0161\u0165astn\u00e9 \u010dasy. Pre mnoh\u00e9 \u00falohy, ktor\u00e9 sme doteraz nevedeli s pomocou umelej…<\/div>\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t\t\t<\/a>\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t<\/article>\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\"\"\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t

Ako mera\u0165 kvalitu vysvetlenia predpovede umelej inteligencie?<\/h3>\n \t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\tsep 26. 2022\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t
Vysvetlite\u013en\u00e1 umel\u00e1 inteligencia: od \u010diernych skriniek k transparentn\u00fdm modelom V minulej \u010dasti seri\u00e1lu sme sa pozreli na to, ak\u00e9 komponenty a vlastnosti by mali ma\u0165 dobr\u00e9 vysvetlenia. Prv\u00fd komponent, zrozumite\u013enos\u0165,…<\/div>\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t\t\t<\/a>\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t<\/article>\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\"\"\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t

S\u00fa vysvetlenia, o ktor\u00fdch si mysl\u00edme, \u017ee s\u00fa dobr\u00e9, naozaj dobr\u00e9?<\/h3>\n \t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\tsep 27. 2023\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t
Vysvetlite\u013en\u00e1 umel\u00e1 inteligencia: od \u010diernych skriniek k transparentn\u00fdm modelom V predo\u0161l\u00fdch \u010dl\u00e1nkoch sme si predstavili tri ve\u013emi zauj\u00edmav\u00e9 koncepty z oblasti vysvetlite\u013enej umelej inteligencie (vysvetlite\u013en\u00e1 AI alebo XAI): V tomto…<\/div>\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t\t\t<\/a>\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t<\/article>\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t\t\t\t\t\t\t\n\t\t\t\t<\/div>\n\t\t\t<\/div>\n\t\t<\/div>\n\t<\/div>\n<\/section>\n<\/div>","protected":false},"excerpt":{"rendered":"

Vysvetlite\u013en\u00e1 umel\u00e1 inteligencia: od \u010diernych skriniek k transparentn\u00fdm modelom M\u00f4\u017ee umel\u00e1 inteligencia rozhodova\u0165 o tom, kto bude prepusten\u00fd z v\u00e4zenia, alebo kto post\u00fapi do \u010fal\u0161ieho kola pracovn\u00fdch pohovorov? M\u00f4\u017ee odpor\u00fa\u010da\u0165…<\/p>\n","protected":false},"author":26,"featured_media":18609,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[349,88,76,67],"tags":[386],"class_list":["post-18618","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-2022-sk","category-pop-science-sk","category-natural-language-processing-sk","category-uncategorized-sk","tag-xai-sk"],"acf":[],"yoast_head":"\nTransparentn\u00e9 modely vs. modely typu \u010dierna skrinka - KInIT<\/title>\n<meta name=\"description\" content=\"Pozrieme sa na dve kateg\u00f3rie modelov umelej inteligencie - transparentn\u00e9 modely a modely typu \u010dierna skrinka. Ak\u00e9 s\u00fa medzi nimi rozdiely a pre\u010do pou\u017e\u00edvame aj netransparentn\u00e9 modely, ke\u010f je schopnos\u0165 pochopi\u0165 ich predpovede tak\u00e1 d\u00f4le\u017eit\u00e1? Ktor\u00e9 modely m\u00f4\u017eeme pova\u017eova\u0165 za transparentn\u00e9 a ktor\u00e9 si, naopak, vy\u017eaduj\u00fa dodato\u010dn\u00fa interpret\u00e1ciu a vysvetlenie ich predpoved\u00ed?\" \/>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/kinit.sk\/sk\/transparentne-modely-vs-modely-typu-cierna-skrinka\/\" \/>\n<meta property=\"og:locale\" content=\"sk_SK\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Transparentn\u00e9 modely vs. modely typu \u010dierna skrinka - KInIT\" \/>\n<meta property=\"og:description\" content=\"Pozrieme sa na dve kateg\u00f3rie modelov umelej inteligencie - transparentn\u00e9 modely a modely typu \u010dierna skrinka. Ak\u00e9 s\u00fa medzi nimi rozdiely a pre\u010do pou\u017e\u00edvame aj netransparentn\u00e9 modely, ke\u010f je schopnos\u0165 pochopi\u0165 ich predpovede tak\u00e1 d\u00f4le\u017eit\u00e1? Ktor\u00e9 modely m\u00f4\u017eeme pova\u017eova\u0165 za transparentn\u00e9 a ktor\u00e9 si, naopak, vy\u017eaduj\u00fa dodato\u010dn\u00fa interpret\u00e1ciu a vysvetlenie ich predpoved\u00ed?\" \/>\n<meta property=\"og:url\" content=\"https:\/\/kinit.sk\/sk\/transparentne-modely-vs-modely-typu-cierna-skrinka\/\" \/>\n<meta property=\"og:site_name\" content=\"KInIT\" \/>\n<meta property=\"article:published_time\" content=\"2022-06-21T11:06:30+00:00\" \/>\n<meta property=\"article:modified_time\" content=\"2023-10-12T12:29:12+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/kinit.sk\/wp-content\/uploads\/2022\/06\/A3.2-Feautre-Web-post-XAI-series.png\" \/>\n\t<meta property=\"og:image:width\" content=\"2401\" \/>\n\t<meta property=\"og:image:height\" content=\"1261\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/png\" \/>\n<meta name=\"author\" content=\"Marianna Palkova\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:creator\" content=\"@kinit\" \/>\n<meta name=\"twitter:site\" content=\"@kinit\" \/>\n<meta name=\"twitter:label1\" content=\"Autor\" \/>\n\t<meta name=\"twitter:data1\" content=\"Marianna Palkova\" \/>\n\t<meta name=\"twitter:label2\" content=\"Predpokladan\u00fd \u010das \u010d\u00edtania\" \/>\n\t<meta name=\"twitter:data2\" content=\"10 min\u00fat\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/kinit.sk\\\/sk\\\/transparentne-modely-vs-modely-typu-cierna-skrinka\\\/#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/kinit.sk\\\/sk\\\/transparentne-modely-vs-modely-typu-cierna-skrinka\\\/\"},\"author\":{\"name\":\"Marianna Palkova\",\"@id\":\"https:\\\/\\\/kinit.sk\\\/#\\\/schema\\\/person\\\/8b175aaaf3267b5bbbbb97e4a6db8cea\"},\"headline\":\"Transparentn\u00e9 modely vs. modely typu \u010dierna skrinka\",\"datePublished\":\"2022-06-21T11:06:30+00:00\",\"dateModified\":\"2023-10-12T12:29:12+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/kinit.sk\\\/sk\\\/transparentne-modely-vs-modely-typu-cierna-skrinka\\\/\"},\"wordCount\":2622,\"image\":{\"@id\":\"https:\\\/\\\/kinit.sk\\\/sk\\\/transparentne-modely-vs-modely-typu-cierna-skrinka\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/kinit.sk\\\/wp-content\\\/uploads\\\/2022\\\/06\\\/A3.2-Feautre-Web-post-XAI-series.png\",\"keywords\":[\"XAI\"],\"articleSection\":[\"2022\",\"Pop science\",\"Natural Language Processing\",\"Uncategorized @sk\"],\"inLanguage\":\"sk-SK\"},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/kinit.sk\\\/sk\\\/transparentne-modely-vs-modely-typu-cierna-skrinka\\\/\",\"url\":\"https:\\\/\\\/kinit.sk\\\/sk\\\/transparentne-modely-vs-modely-typu-cierna-skrinka\\\/\",\"name\":\"Transparentn\u00e9 modely vs. modely typu \u010dierna skrinka - KInIT\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/kinit.sk\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/kinit.sk\\\/sk\\\/transparentne-modely-vs-modely-typu-cierna-skrinka\\\/#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/kinit.sk\\\/sk\\\/transparentne-modely-vs-modely-typu-cierna-skrinka\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/kinit.sk\\\/wp-content\\\/uploads\\\/2022\\\/06\\\/A3.2-Feautre-Web-post-XAI-series.png\",\"datePublished\":\"2022-06-21T11:06:30+00:00\",\"dateModified\":\"2023-10-12T12:29:12+00:00\",\"author\":{\"@id\":\"https:\\\/\\\/kinit.sk\\\/#\\\/schema\\\/person\\\/8b175aaaf3267b5bbbbb97e4a6db8cea\"},\"description\":\"Pozrieme sa na dve kateg\u00f3rie modelov umelej inteligencie - transparentn\u00e9 modely a modely typu \u010dierna skrinka. Ak\u00e9 s\u00fa medzi nimi rozdiely a pre\u010do pou\u017e\u00edvame aj netransparentn\u00e9 modely, ke\u010f je schopnos\u0165 pochopi\u0165 ich predpovede tak\u00e1 d\u00f4le\u017eit\u00e1? Ktor\u00e9 modely m\u00f4\u017eeme pova\u017eova\u0165 za transparentn\u00e9 a ktor\u00e9 si, naopak, vy\u017eaduj\u00fa dodato\u010dn\u00fa interpret\u00e1ciu a vysvetlenie ich predpoved\u00ed?\",\"breadcrumb\":{\"@id\":\"https:\\\/\\\/kinit.sk\\\/sk\\\/transparentne-modely-vs-modely-typu-cierna-skrinka\\\/#breadcrumb\"},\"inLanguage\":\"sk-SK\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/kinit.sk\\\/sk\\\/transparentne-modely-vs-modely-typu-cierna-skrinka\\\/\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"sk-SK\",\"@id\":\"https:\\\/\\\/kinit.sk\\\/sk\\\/transparentne-modely-vs-modely-typu-cierna-skrinka\\\/#primaryimage\",\"url\":\"https:\\\/\\\/kinit.sk\\\/wp-content\\\/uploads\\\/2022\\\/06\\\/A3.2-Feautre-Web-post-XAI-series.png\",\"contentUrl\":\"https:\\\/\\\/kinit.sk\\\/wp-content\\\/uploads\\\/2022\\\/06\\\/A3.2-Feautre-Web-post-XAI-series.png\",\"width\":2401,\"height\":1261},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/kinit.sk\\\/sk\\\/transparentne-modely-vs-modely-typu-cierna-skrinka\\\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Home\",\"item\":\"https:\\\/\\\/kinit.sk\\\/sk\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"2022\",\"item\":\"https:\\\/\\\/kinit.sk\\\/sk\\\/category\\\/2022-sk\\\/\"},{\"@type\":\"ListItem\",\"position\":3,\"name\":\"Transparentn\u00e9 modely vs. modely typu \u010dierna skrinka\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/kinit.sk\\\/#website\",\"url\":\"https:\\\/\\\/kinit.sk\\\/\",\"name\":\"KInIT\",\"description\":\"Vyu\u017e\u00edvame v\u00fdskum pre \u013eud\u00ed a priemysel\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/kinit.sk\\\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"sk-SK\"},{\"@type\":\"Person\",\"@id\":\"https:\\\/\\\/kinit.sk\\\/#\\\/schema\\\/person\\\/8b175aaaf3267b5bbbbb97e4a6db8cea\",\"name\":\"Marianna Palkova\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"Transparentn\u00e9 modely vs. modely typu \u010dierna skrinka - KInIT","description":"Pozrieme sa na dve kateg\u00f3rie modelov umelej inteligencie - transparentn\u00e9 modely a modely typu \u010dierna skrinka. Ak\u00e9 s\u00fa medzi nimi rozdiely a pre\u010do pou\u017e\u00edvame aj netransparentn\u00e9 modely, ke\u010f je schopnos\u0165 pochopi\u0165 ich predpovede tak\u00e1 d\u00f4le\u017eit\u00e1? Ktor\u00e9 modely m\u00f4\u017eeme pova\u017eova\u0165 za transparentn\u00e9 a ktor\u00e9 si, naopak, vy\u017eaduj\u00fa dodato\u010dn\u00fa interpret\u00e1ciu a vysvetlenie ich predpoved\u00ed?","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/kinit.sk\/sk\/transparentne-modely-vs-modely-typu-cierna-skrinka\/","og_locale":"sk_SK","og_type":"article","og_title":"Transparentn\u00e9 modely vs. modely typu \u010dierna skrinka - KInIT","og_description":"Pozrieme sa na dve kateg\u00f3rie modelov umelej inteligencie - transparentn\u00e9 modely a modely typu \u010dierna skrinka. Ak\u00e9 s\u00fa medzi nimi rozdiely a pre\u010do pou\u017e\u00edvame aj netransparentn\u00e9 modely, ke\u010f je schopnos\u0165 pochopi\u0165 ich predpovede tak\u00e1 d\u00f4le\u017eit\u00e1? Ktor\u00e9 modely m\u00f4\u017eeme pova\u017eova\u0165 za transparentn\u00e9 a ktor\u00e9 si, naopak, vy\u017eaduj\u00fa dodato\u010dn\u00fa interpret\u00e1ciu a vysvetlenie ich predpoved\u00ed?","og_url":"https:\/\/kinit.sk\/sk\/transparentne-modely-vs-modely-typu-cierna-skrinka\/","og_site_name":"KInIT","article_published_time":"2022-06-21T11:06:30+00:00","article_modified_time":"2023-10-12T12:29:12+00:00","og_image":[{"width":2401,"height":1261,"url":"https:\/\/kinit.sk\/wp-content\/uploads\/2022\/06\/A3.2-Feautre-Web-post-XAI-series.png","type":"image\/png"}],"author":"Marianna Palkova","twitter_card":"summary_large_image","twitter_creator":"@kinit","twitter_site":"@kinit","twitter_misc":{"Autor":"Marianna Palkova","Predpokladan\u00fd \u010das \u010d\u00edtania":"10 min\u00fat"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/kinit.sk\/sk\/transparentne-modely-vs-modely-typu-cierna-skrinka\/#article","isPartOf":{"@id":"https:\/\/kinit.sk\/sk\/transparentne-modely-vs-modely-typu-cierna-skrinka\/"},"author":{"name":"Marianna Palkova","@id":"https:\/\/kinit.sk\/#\/schema\/person\/8b175aaaf3267b5bbbbb97e4a6db8cea"},"headline":"Transparentn\u00e9 modely vs. modely typu \u010dierna skrinka","datePublished":"2022-06-21T11:06:30+00:00","dateModified":"2023-10-12T12:29:12+00:00","mainEntityOfPage":{"@id":"https:\/\/kinit.sk\/sk\/transparentne-modely-vs-modely-typu-cierna-skrinka\/"},"wordCount":2622,"image":{"@id":"https:\/\/kinit.sk\/sk\/transparentne-modely-vs-modely-typu-cierna-skrinka\/#primaryimage"},"thumbnailUrl":"https:\/\/kinit.sk\/wp-content\/uploads\/2022\/06\/A3.2-Feautre-Web-post-XAI-series.png","keywords":["XAI"],"articleSection":["2022","Pop science","Natural Language Processing","Uncategorized @sk"],"inLanguage":"sk-SK"},{"@type":"WebPage","@id":"https:\/\/kinit.sk\/sk\/transparentne-modely-vs-modely-typu-cierna-skrinka\/","url":"https:\/\/kinit.sk\/sk\/transparentne-modely-vs-modely-typu-cierna-skrinka\/","name":"Transparentn\u00e9 modely vs. modely typu \u010dierna skrinka - KInIT","isPartOf":{"@id":"https:\/\/kinit.sk\/#website"},"primaryImageOfPage":{"@id":"https:\/\/kinit.sk\/sk\/transparentne-modely-vs-modely-typu-cierna-skrinka\/#primaryimage"},"image":{"@id":"https:\/\/kinit.sk\/sk\/transparentne-modely-vs-modely-typu-cierna-skrinka\/#primaryimage"},"thumbnailUrl":"https:\/\/kinit.sk\/wp-content\/uploads\/2022\/06\/A3.2-Feautre-Web-post-XAI-series.png","datePublished":"2022-06-21T11:06:30+00:00","dateModified":"2023-10-12T12:29:12+00:00","author":{"@id":"https:\/\/kinit.sk\/#\/schema\/person\/8b175aaaf3267b5bbbbb97e4a6db8cea"},"description":"Pozrieme sa na dve kateg\u00f3rie modelov umelej inteligencie - transparentn\u00e9 modely a modely typu \u010dierna skrinka. Ak\u00e9 s\u00fa medzi nimi rozdiely a pre\u010do pou\u017e\u00edvame aj netransparentn\u00e9 modely, ke\u010f je schopnos\u0165 pochopi\u0165 ich predpovede tak\u00e1 d\u00f4le\u017eit\u00e1? Ktor\u00e9 modely m\u00f4\u017eeme pova\u017eova\u0165 za transparentn\u00e9 a ktor\u00e9 si, naopak, vy\u017eaduj\u00fa dodato\u010dn\u00fa interpret\u00e1ciu a vysvetlenie ich predpoved\u00ed?","breadcrumb":{"@id":"https:\/\/kinit.sk\/sk\/transparentne-modely-vs-modely-typu-cierna-skrinka\/#breadcrumb"},"inLanguage":"sk-SK","potentialAction":[{"@type":"ReadAction","target":["https:\/\/kinit.sk\/sk\/transparentne-modely-vs-modely-typu-cierna-skrinka\/"]}]},{"@type":"ImageObject","inLanguage":"sk-SK","@id":"https:\/\/kinit.sk\/sk\/transparentne-modely-vs-modely-typu-cierna-skrinka\/#primaryimage","url":"https:\/\/kinit.sk\/wp-content\/uploads\/2022\/06\/A3.2-Feautre-Web-post-XAI-series.png","contentUrl":"https:\/\/kinit.sk\/wp-content\/uploads\/2022\/06\/A3.2-Feautre-Web-post-XAI-series.png","width":2401,"height":1261},{"@type":"BreadcrumbList","@id":"https:\/\/kinit.sk\/sk\/transparentne-modely-vs-modely-typu-cierna-skrinka\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https:\/\/kinit.sk\/sk\/"},{"@type":"ListItem","position":2,"name":"2022","item":"https:\/\/kinit.sk\/sk\/category\/2022-sk\/"},{"@type":"ListItem","position":3,"name":"Transparentn\u00e9 modely vs. modely typu \u010dierna skrinka"}]},{"@type":"WebSite","@id":"https:\/\/kinit.sk\/#website","url":"https:\/\/kinit.sk\/","name":"KInIT","description":"Vyu\u017e\u00edvame v\u00fdskum pre \u013eud\u00ed a priemysel","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/kinit.sk\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"sk-SK"},{"@type":"Person","@id":"https:\/\/kinit.sk\/#\/schema\/person\/8b175aaaf3267b5bbbbb97e4a6db8cea","name":"Marianna Palkova"}]}},"_links":{"self":[{"href":"https:\/\/kinit.sk\/sk\/wp-json\/wp\/v2\/posts\/18618","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/kinit.sk\/sk\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/kinit.sk\/sk\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/kinit.sk\/sk\/wp-json\/wp\/v2\/users\/26"}],"replies":[{"embeddable":true,"href":"https:\/\/kinit.sk\/sk\/wp-json\/wp\/v2\/comments?post=18618"}],"version-history":[{"count":14,"href":"https:\/\/kinit.sk\/sk\/wp-json\/wp\/v2\/posts\/18618\/revisions"}],"predecessor-version":[{"id":28796,"href":"https:\/\/kinit.sk\/sk\/wp-json\/wp\/v2\/posts\/18618\/revisions\/28796"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/kinit.sk\/sk\/wp-json\/wp\/v2\/media\/18609"}],"wp:attachment":[{"href":"https:\/\/kinit.sk\/sk\/wp-json\/wp\/v2\/media?parent=18618"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/kinit.sk\/sk\/wp-json\/wp\/v2\/categories?post=18618"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/kinit.sk\/sk\/wp-json\/wp\/v2\/tags?post=18618"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}