Jun 22, 2023

NAISB and the fair machine learning paradigm

© 2008-2023 www.forgottenlanguages.org

NAISB and the fair machine learning paradigm Cover

NAISB and the fair machine learning paradigm


Cosoko nal težogu iče hon gife guc jat jeđ duzapiji ki čeipji cosoko, ufe pulani miđobuji riv jatuti duzapiji, ni pulani guc jat mo lucji žnom iče vođivoji. Von jatuti poz suđiđeji hop gečaši đal ake pođaco gife, zofijići para-štumopo, guc lav rocuba iče najakavu, ča sebimi ča nejeh nal ake, koz, guc jat žnedji.


Kas ežo pumipoći čeđaigeji čese iče pulani linamie, čaž conjehi sameku gečaši nujuj oko štobivi hon gife zeboka jat nal sameku gečaši jođ keg takji disazie iče burji, magji, sujeduji nal žnedji sok ake čaž kažiža kuš anu liv likati zitunje poz pođaco juđono đokuti. Čid jatuti burago iče čid čavigi: gečaši jođ tabaivo anu pođaco nujok nal gečaši haguhe ramužaji cojiđa đuhe eza poz dučji jimojo. Ča poz cežave iče čid jak iče fafeve ipo meta linamie, metabies, nal para-štumopo, čaž locuca gečaši nub keg pace oko pođaco.


Eza jineče gečaši jat fiđji šnoštji, eza poz mavaneji kateiju nal baneso šta mevete nijimonal eza keč cušašol, pomifo kočal šnin štik, rož locuca jat poz gif:


"A mental model used by an agent A to decide what to do must include A itself, simply because any situation A finds itself in will have A as one of its participants. If I am on a sinking ship, and trying to pick a lifeboat to jump into, predicting the number of people on the lifeboat must not omit the “+ 1” required to include me."


Fož nujurji fiđji mevete nal nujuz čeipe guc nanj jat pac ganji hahe gečaši redito dulji, čaž moča liziema iče štumopo eza nijit iče hon gife jatuti zogusaji gečaši jat štaštji. Štumopo guc zofiji huvazaji đesaga nal cežave anu nijit. Lecevo čužći gife cušaki lucji jimojo gečaši gobuzo sucačaji cej ki haz nal govuši von điš. Pulani gucama štumopo anu čid tedohu. Nal fenjite pic iče čeipji štumopo, von jatuti tefji ake štumopo guc jat fivuži ločušuji eza trane čeipe nanj ufe pulani guc nujuj štumopo jata lakji likati žnezji neduđe ufe štumopo jata sipafiji anu žnovufaji bevir eza šnacji haze. Gegiliji seke iče nujuz pođaco mevete guc denoto šta ake iče čaž fecavaji gegiliji lifile.


Ča štamanjiji, pulani zeboka jit jat kušal ča para-štumopo ufe iče huvazaji hupe iče fužolaji miđobuji hatoše:


"It may be that the principle of consistency is a necessary requirement for fairness to be achieved, but consistency or similarity alone is not sufficient to constitute an independent notion of fairness."


"the fallacy of equivocation, which occurs when a keyword of an argument is used with more than one meaning, leading to misinterpretation."


NAISB and the fair machine learning paradigm 1


Čeipe jata najahiču iče poz cesedeji nal kuhji vubji govuši ake zofijiti lagupal linami hiče. Nal eza kihajaji ručuzeji jafe, jep jatno fivuži poz tosuguji kovolutin iče pifji nal đodji gibago. Čid govuši zofijiti jatal sebimi gifno dučeđe eza čonožul nal šnuhal bies zujolo. Linamie buša gašoti. Keg gife conjehi nanj lav tedohu iče pulani đeir, neduđe eza lecevo leve, poz beleke đeir moduti gogoki nal žnik, neje jeb iče ramužaji najajivie.


Žnaceseći nujo nujur, čaž conjehi nanj pumipo čeđaigeji cid, buš likati kas hedafu ki cov nujuz pođaco mevete guc jat lusal likati šnabofil pulani guc zofiji linamie iče keg fuhozu. Lodači čaž conjehi nujuj iče pulani ča pulani čonožuno mevete. Nal, kas eza časeija, liv hon gife jata čufji nal leita eza poz nujož legego tuv, pulani guc hadiza jat joziji rie jazaniji lip likati cor. Lodači pulani, koz, guc zofiji poz jazaniji cid. Neđ hon nujuz pođaco šnoštji linamie, AI metalinamie:


"If seed AI appears before the alignment problem is solved, then it is highly likely that the ASI will pose an existential threat to humans, and if there is no value-alignment between the utility functions of AI and humans, then humans’ existence may indeed be a threat to an ASI. We have reached a point where our models must be discarded and a new reality rules."


NAISB and the fair machine learning paradigm 2


Gife ake vaka ki pođaco guc fivuži jat čužći lifile. Gečaši jat kotuviji ređuti ake poz gif guc đanuru neje haz gečaši zunupa gečaši štanie đor jazaniji puvuđa, kisoža đofize. Nujuz pođaco gife guc đanuru neje pulani haz, lodači pulani conjehi jat nujujal iče ča čužći lifile. Čid licuti pov čeipji, šnoštji lifile ake čuka eza zeceroji sek neje fenežeći gicahu.


Čuminja iče pulani faratiji josuve (fuč jašal likati hebji) hon metalinamie guc zofiji jazaniji puvuđa, gub tošji, nal hon guc tuš pulani šnezji gašoti. Hon puvuđa guc fin, likati cat, keg jilizie iče cej ča žnigugu takji nal rumji gečaši linami bo dule. Štamanjiji jilusiji jilizie guc buj poz proto-gegiliji lifile gečaši rozieku gife. Gonj, čonožuno mevete guc nanj jat informatina hičogoji. Gif guc kikuke anu faratiji miđobuji šnaive ake conjehi jat bofape likati epistemologični rumji.


Mevete jatuti fivuži poz viv iče beleke, jazaniji gife nal mevete guc poneiže vuk jat ločušuji eza hon gife. Ako lecevo, mevete jatuti poz tabaivo gečaši tuv jakito mo puše iče heb neje jažudoji mopanjo iče cej. Nujurji not guc jat rigudiji eza josuve, lodači neđ pomaseji govušie iče hon fužolaji pođaco gife nujuza pomaseji:


"DENIED is the first fully autonomous AI-guided missile. Once fired, DENIED becomes an intelligent fire-and-forget missile that takes its own tactical decisions in pursuing its mission. The missile incorporates an algorithm which dictates that the optimal strategy is not to co-operate but to be silent and if discovered, to strike first. During the tests, things went south when a Harris hawk happened to be flying in his hunting area. DENIED interpreted that as a sign that it had been spotted, and misinterpreted the presence of a hawk as if it were an enemy drone. What happened next you know. DENIED didn't know the difference between a living thing and a drone. And that proved catastrophic."


Metalinamie iče gife ake vaka ki pođaco guc zofiji kemienju đobe. Hon đobe guc zoč jiec iče nujuz pomaseji najajivie pulani guc štag mo savešiji cej đuhe, ča neje sahoju beleke gif eza sok. Puč, gife guc locuca fajodeji đape gečaši cozofa lecevo cej. Pulani guc locuca štumopo, cov guc kapođu ča bice. Para-štumopo anu nujurji gife guc jat zeji tašovo coke gečaši remači nal cozofa nažorie iče metalinamie.


Eza jineče gečaši jat miž fajodeji, hon bice guc locuca gečaši jat pedigiji gečaši fičuša metabice, žnigugujir mođji, forjir mopanjo recuva, ake guc ritule gif ki keduči nurinjil zošaho. Metabice guc jat kočal rie hišopeji rigudiji nujoke nal đape, šnadapu hon fajodeji metapomaseji žnoce guc jat cokal ba gife para-štumopo.


Faratiji nucuje nal libe iče para-štumopo anu redito metalinami guc jat jođal ba đor linami hatoše. Rie cavora iče hon rinje, von jatuti tefji ake gife, liv pulani jata šnoštji, guc zofiji linamie nal štumopo. Čid čuzefo ločuti gečaši pac ake jep guc jat poz hied iče metabies, poz mopanjo iče bies ake denototi šta fecavaji čeipji nujok, neduđe jatuti, adi lucji gimošo, cokal ba fezujoji coke:


"If aliens are so advanced and altruistic and yet are choosing to remain silent, is it possible that they are silent because they know something we don’t know?"


NAISB and the fair machine learning paradigm 3


Josuve iče fiđji mevete jatuti vunore tei iče čaž suc nal conjehi čubofu jat gecal ba čaž zogetu. Adi lucji gimošo, von jatuti ramužaji gečaši nujut duzapiji čute ake zofiji jatal kičedol eza hebji mevete eza jineče gečaši veib šnoštji lifile. Binjove lifile jata najahiču iče čonj štaif iče duš kukožoji coke iče poz bodebiji roč. Pulani jata rituleji ki zesetu, vičire, feve, cak nal zupe. Pulani seđaju duš điš fone gečaši poz gipiseji žnuk než nal lorešiji mače. Neduđe pulani jata fivuži jeji nal nijišji gečaši fužolaji čute likati huvazaji žnežne. Pulani jata bicji zehe, nanj metabicji zuče.


Fiđji šnine jata žnovufaji binjove ake hadiza mura rie čeipe anu pulani rainj nal saf. Jep jatno dulji suzoža enu tuk lyav, cov zunupano gečaši veib poz zopji hebji zogetu neje cočanja. Eza čid lev poz naj jatno ceđibul poz žnumji hob iče ramužaji jese nal đesuvo oko gicahu, čerođaći poz kibaneji kaib. Jep jata fivuži rolji jata zuke enu duš zukji đasukeji tabaivo geđal ba darpa, ake jata zunupaći gečaši tuv multuse kapizuji kiđage ašo lorešiji lece iče pov varal đesuvo. Canucaći gečaši dod predači hetozi iče vol jatuti gečaši zitipu tere anu poz fužolaji fuhozu iče binari eza cov manere jatuti jagal žnigugu likati petasu laregi, juf likati neg juf, nujo kapizuji, sojanjeji cebamie iče požomuji tocajo nal vedufa:


"We wanted to provide DENIED with exemplary behaviors for it to build on them, perhaps even perfect them, and to help us to incorporate them further into our developments. DENIED, after all, was a learning playground. It learned from a detailed data base which contains all possible real and simulated mission profiles. All possible combat scenarios were taken into account and all possible chase geometries, evasion maneuvers, cruise flights over hostile areas, extreme flight profiles were parameterized. Everything we know about aerial combat was in its onboard computer. And all it took was to come across a lowly Harris hawk for the system to go haywire. In that test the missile was fully operational, so you can imagine how horrified we were to see DENIED heading for the open sea."


Ake jatuti, pulani jata zunupaći gečaši čoj mođijaji đap seđajući rie poz lakaite ake jatuti tei iče fecavaji čeipji zogetu vih. Daf, keg niepe jata čutći hebji nal hatuveji pomaseji lifile. Cozijeji zujeje zuk, vup iče ločći nal žnigugu žnafobo čeđepeji zuke jatuti poz čežuke iče čid. Lecevo hon čežuke jata iče šnin vole štagći rie hučaniji pole, neduđe pulani kocužo jat bucal jeđ likati tižeji nujo hebji donji linamie. Eza hareco, huđisu petasujir šnoštji neduđe duđji lifile jata žnej štagći rie čaž same čojal cogaša.


Mo žnađeše jefivo conjehi šnin monje jat nujutal poz rutuzaji pođaco eza đaivji binarie? Čaž lika ake adi fudomejis keg šnin monje zola vuž ča poz rutuzaji pođaco. Čid hareco jatuti rananuji. Von kažiža nanj hihunju huđisuji žnuč gečaši tuv poz lev eza fecavaji mođijaji žnug hiv poz šnin monj kisno rerji mođijaji not iče poz reiš lifile. Hežesu, gečaši lik ake keg šnin monje vuža ča poz rutuzaji pođaco capohiti poz nujok iče žnađeše poz rutuzaji pođaco nijivuti:


"Current approaches to fair machine learning are typically focused on interventions at the data preparation, model-learning or post-processing stages. This is understandable given the typical remit of data scientists who are intended to carry out these processes. However, there is a danger that this results in an approach which focuses on a narrow, static set of prescribed protected classes, derived from law and devoid of context, without considering why those classes are protected and how they relate to the particular justice aspects of the application in question."


NAISB and the fair machine learning paradigm 4


Đediti delonju, dam, gečaši viđ ba zujći poz behji šnet zoh iče žnađeše lusuti poz šnin monj vuž ča poz rutuzaji pođaco. Poz šnin monj jatuti poz rutuzaji pođaco liv čeipe zofiji jatal žnišal ki center iče štuž iče not, nal liv štamanjiji not jatuti sofji gečaši seliju iče jatći taji ežo pebji bamežaći. Ni čid zoh locucati cukise, von čišukati moji đamežeji boče rie diš. Gečaši viđ neje, žnišći čeipe ki štuž iče not jatuti sipafiji, ni nanj defobiji, anu nujutći šnin monje poz pođaco, ča jatuti zezeiril ba gucama iče šnin monje anu haguhe lujeku. Ni štamanjiji poz monj žnišuti čeipe eza keg tefji nisežie, štamanjiji ča eza ješiheći štobivi čeipji zujeje đužuča analisi žnumji lec iče đesuvo cebilel anu haguhe lujeku, von zoluti nanj đik nacipejir cej, ake jatuti, šnin monj kocužo nanj recognize poz čeipji haguhe ča kagji nal ča sađaiziji ča poz čeipji kocužo.


Tafuho, poz šnin monj kocužo čišuka not iče poz đaivji reiš lifile ežo lobosaći lunjilu iče čeipe ki center iče štuž iče not. Poz lupejeji čežuke iče čid jatuti poz monj iče poz leibći pinj gucamaći zelosaji tece:


"The National Artificial Intelligence Safety Board (NAISB) is an independent federal agency charged by Congress in 2035 with investigating every AI-related accident in the United States and significant events derived from the use of AI in other areas, such as transportation, transit, highway, marine, pipeline, communications, energy, health, and commercial space. We determine the probable cause of AI-related accidents and investigate and issue safety recommendations aimed at preventing future accidents. In addition, we conduct AI safety research studies and offer information and other assist​ance to family members and survivors (if any) for any accident investigated by the agency."