- Ukuhunyushwa kwemishini kuhlola izisindo, ukusebenza, kanye nezifunda zangaphakathi ukuchaza ukuthi amanethiwekhi e-neural kanye nama-LLM enza kanjani izibalo zawo.
- Amamodeli ahlela izincazelo zibe yizikhala zomqondo ezisezingeni eliphezulu, kanye nemibono emelelwa njengezikhombisi-ndlela eziqondile kuma-vector.
- Amathuluzi anjenge-"microscopes" kanye nama-sparse autoencoders akuvumela ukuthi ukhiphe, uhlaziye, futhi uphathe izici zangaphakathi zamamodeli.
- Izicelo ezifana nokuhunyushwa kwe-geospatial zibonisa ukuthi ama-LLM ahlela kanjani ulwazi lwendawo, okusondeza i-AI ezingxoxweni mayelana nokuqonda nokuphepha.
Ukuhumusha ngobuchwepheshe sekuba ngomunye wemigqa yocwaningo ethokozisayo nebaluleke kakhulu ngaphakathi kwe-AI yanamuhla.Lokhu kubaluleke kakhulu njengoba amanethiwekhi ajulile e-neural kanye namaModeli Olimi Olukhulu (ama-LLM) eqala ukuthonya izinqumo cishe kuzo zonke izinkambu. Esikhundleni sokubheka nje ukusebenza kokugcina kwemodeli, le ndlela iyabuza: yini ngempela eyenzekayo ngaphakathi, ezisindweni nasekusebenzeni, lapho i-AI yenza isibikezelo, ibhala umbhalo, noma ixazulula inkinga eyinkimbinkimbi?
Igama elithi "ibhokisi elimnyama" alikaze libe nomthelela omkhulu kunamanje.Amakhulu ezigidi zabantu asebenzisa ama-chatbot nsuku zonke, kodwa ngisho namaqembu athuthukisa lezi zinhlelo awaqondi ngokugcwele ukuthi afinyelela kanjani ezimpendulweni ezithile, ukuthi kungani "eqamba amanga" amaqiniso, noma ukuthi angaziphatha kanjani ngokukhohlisa. Ukuchazwa kwezindlela kuvela ngokunembile ukuze kuvule leli bhokisi elimnyama, kudwetshwe izindlela zalo zangaphakathi, futhi kuxhunywe ama-neurons, izinsiza, kanye nezifunda emiqondweni esingayiqonda.
Kuyini ngempela ukuhunyushwa kwemishini?
Ukuhunyushwa kwemishini isifundo esihlelekile sesakhiwo sangaphakathi samamodeli e-AI, sigxile ezisindweni, ekusebenzeni, kanye "nokubala" okuphakathi.ukuqonda ukuthi benza kanjani imisebenzi yabo. Esikhundleni sokuphatha inethiwekhi yemizwa njengebhulokhi elingacacile eliguqula okokufaka kube umphumela, le ndawo izama ukuhlakaza imodeli ibe yizingxenye ezincane—ama-neurons, amakhanda okunaka, izendlalelo, izici eziqondile—futhi ixhumanise ingxenye ngayinye nokuziphatha okubonakalayo.
Inhloso eyinhloko akukhona nje "ukuchaza kamuva" isinqumo esihlukile, kodwa ukwakha imephu eningiliziwe yokubala kwangaphakathi kwemodeli.Lokhu kuhilela ukuhlonza ukuthi yimaphi ama-neuron noma inhlanganisela yama-neuron amelela amaphethini athile (njengamagama afanele, izakhiwo zekhodi, amathoni emizwa, imiyalelo enonya), ukuthi lokhu kumelela kuhlanganiswa kanjani kuzo zonke izendlalelo, nokuthi konke lokhu kuphumela kanjani emphumeleni othize.
Lo mbono ubulokhu ukhula ngokushesha emphakathini wesayensi.Njengoba kunemihlangano yokusebenzela ezinikele (njenge-workshop yokuqala enkulu mayelana nokuhumusha kwemishini emihlanganweni emikhulu yokufunda komshini), izinkampani eziningi ezintsha ezigxile esihlokweni, kanye nenani elikhulayo lamathuluzi okuhlaziya, inani lamaphepha athunyelwe kuma-workshop akhethekile lidlula kalula ikhulu ngohlelo ngalunye, okubonisa ukuthi lokhu akusekho ndawo futhi sekuphenduke insimu ehlanganisiwe ekwandisweni okugcwele.
Inselele enkulu ukunciphisa igebe phakathi kokusebenza okumangalisayo kwamamodeli nokuqonda kwethu ngawo.Uma nje siqhubeka nokuphatha ama-LLM kanye namanethiwekhi ezinzwa njengezimfihlakalo zezibalo, kuzoba nzima kakhulu ukubikezela ukuziphatha okunqenqemeni, ukuhlonza ubuthakathaka obuyinkimbinkimbi, ukubona ukuxhashazwa, nokusebenzisa lezi zinhlelo ngokwethembeka ezimweni ezibucayi.
Izikhala Zomqondo kanye Ne-Hypothesis Yokumelwa Okuqondile
Enye yezindlela ezinamandla kakhulu zokuqonda ukuhunyushwa kwemishini umqondo wokuthi amanethiwekhi e-neural akha "izikhala zomqondo" ezinobukhulu obuphezulu.Esikhundleni sokucabanga ngezincazelo njengezincazelo kusichazamazwi, singazibona njengezici esikhaleni esikhulu sevektha, ezifihliwe kunethiwekhi, ezakhiwe yizisindo kanye nokusebenza kuzo zonke izendlalelo.
Lesi sikhala asisona esibonakalayo; siwumphumela ongemuhle wendlela inethiwekhi ecubungula ngayo amasignali.Okokufaka ngakunye (imiqondo yombhalo efana negama, iphikseli, umsindo, igama lendawo, isiqeshana sekhodi) kuhlotshaniswa nevektha esikhaleni esinobukhulu obuningi. Le vektha ithwebula konke imodeli "ebhekwa njengokufanelekile" mayelana nalokho okufakwayo, ngokusekelwe ekuqeqeshweni kwayo, futhi ingabhala ama-nuances e-semantic, isitayela, umongo, inhloso, nokunye okuningi.
I-Hypothesis ebizwa ngokuthi i-Linear Representation Hypothesis ithi eminingi yale mibono yangaphakathi ingabhekwa njengeziqondiso eziqondile kulesi sikhala.Ngamanye amazwi, kukhona uhlangothi olulodwa oluhambisana "nokudumisa," olunye oluya "ephutheni lokubhala ikhodi," olunye oluya "emuva kwedijithali," njalo njalo. Imiqondo eyinkimbinkimbi kakhulu ingakhiwa ngokuhlanganisa eziningana zalezi ziqondiso eziyisisekelo.
Lokhu kusho ukuthi noma yiluphi uhlobo lolwazi — ulimi, umbono, umsindo, ukunyakaza — lungamelwa njengezinto eziveza ulwazi kulesi sikhala esifanayo somqondo.Uma i-LLM icubungula umusho, isibonelo, empeleni ilandela indlela kuleso sikhala, ibuyekeza i-context vector ngethokheni ngayinye ukuze ibambe incazelo eqoqwe kuze kube yileso sikhathi.
Lo mbono uchaza nokuthi kungani kungenzeka "ukuzulazula" phakathi kwemibono, ukuyihlanganisa noma ukuyisusa.Ngokuhambisa i-vector kusuka endaweni eyodwa kuya kwenye ngendlela ethile, singasuka ku-"cat" siye ku-"fat cat," "smart cat," "lazy cat"; noma ngisho nokushintsha phakathi kwezilimi, sigcina umqondo ofanayo oyisisekelo ngenkathi ubuso (igama) bushintsha.
Imibono echazwa ngomehluko: akukho lutho olukhona oluzimele.
Isici esithakazelisayo salolu hlobo ukuthi, kunethiwekhi, akukho lutho olunencazelo ephelele; konke kuchazwa ubudlelwano bayo nenye indawo.Umqondo "wekati" awuveli encazelweni yombhalo yangaphakathi, kodwa uvela endaweni yawo maqondana "nendlovu," "itafula," "inja," "obomvu," "onoboya," "olula," "osindayo," njalo njalo.
Uma wazi ukuthi indlovu inkulu futhi inzima kunekati, uboya obuncane, obunokwakheka okuhlukile, nokuthi itafula likhanya kakhulu kunezimbili, hhayi elinoboya, likhulu kunekati futhi lincane kunendlovu.Bese kuqala ukuvela isakhiwo: "usayizi," "isisindo," "ukuthungwa," "izinwele," "ukukhanya." Lezi zilinganiso azidingi ukuhambisana ngqo nalezo esizisebenzisa ngomqondo ojwayelekile, kodwa zisebenza njengezikhonkwane ezihlela imiqondo ngendlela ewusizo kumodeli.
Njengoba isikhala sigcwala imiqondo, lobu budlelwano obuhlanganisayo bucwengisisa imiqondo ngokwayo kanye "nobukhulu obucashile."Ngamazwi angokoqobo, lapho imodeli ifunda futhi ilungisa izisindo zayo, kulapho lezi zindlela zangaphakathi zikhula khona, okuvumela izibikezelo ezicashile nezifaneleka ngokomongo.
Kubalulekile ukukhumbula ukuthi "usayizi," "isisindo," noma "uboya" ziyizingathekiso ezilula.Eqinisweni, ubukhulu obusetshenziswa yi-AI bungabamba amaphethini ayinkimbinkimbi kakhulu angangeni ezigabeni ezilula zabantu. Kungaba yinhlanganisela engeyona into encane yezici zokubhala, ezesisho, ezibonwayo, ezisendaweni, ezingokwesitayela, nezinye.
Ngomqondo othile, lesi sikhala sevektha sakha "imodeli yomhlaba" yangaphakathi.Akuyona nje into engaqondakali: kuyinto engokoqobo eyenzeka namuhla kumanethiwekhi e-neural kanye nama-LLM. Uma sithi imodeli "iyaqonda" okuthile, lokho esikubonayo empeleni kuwumphumela waleyo nqubo yokubeka kanye nokuhlobana kwama-vector kuleso sikhala somqondo esingacacile.
Kusukela ku-microscopy yezinsizakusebenza kuya ezinkampanini ezinkulu ze-AI
Eminyakeni yamuva nje, ukuhunyushwa kwemishini kuye kwathuthuka ngenxa yamathuluzi amasha asebenza, ngokomfanekiso, njenge-microscope yamamodeli olimi.Esikhundleni sokubuka nje okufakwayo kanye nemiphumela, abacwaningi baqala ukuhlola ngqo ukusebenza kwangaphakathi kanye nezindawo ezithile zesikhala se-vector lapho kuhlala khona imiqondo ethile.
Izinkampani ezifana ne-Anthropic, i-OpenAI, i-Google DeepMind, kanye namaphrojekthi afana ne-Neuronpedia bezilokhu zihola lo mzamo.Isibonelo, i-Anthropic yamemezela inqubo ebizwa ngokuthi "i-microscope" yokubheka ngaphakathi kwemodeli yayo iClaude nokuhlonza izici zangaphakathi ezihambisana nemibono eqashelwayo, njengoMichael Jordan, iGolden Gate Bridge, noma ngisho nemibono engaqondakali efana "nokuthopha" kanye "neminyango yangemuva yedijithali."
Ngemva kwalokho, ucwaningo lwaqhubekela phambili ekulandeleni yonke imithombo.Lokhu akubonisi nje kuphela ukuthi i-neuron noma i-vector kuhlotshaniswa nomqondo, kodwa futhi nendlela lowo mqondo osebenza ngayo, oguqulwa ngayo, futhi ohlanganiswa ngayo kuzo zonke izendlalelo, kusukela emyalweni wokuqala kuya empendulweni yokugcina. Lokhu kusenza, isibonelo, siqonde ukuthi yiziphi izingxenye zemodeli ezihlanganyela ekuziphatheni okuthile okukhohlisayo noma ekucabangeni okungaqondakali.
Amaqembu avela ku-OpenAI kanye ne-Google DeepMind aqale ukusebenzisa amasu afanayo ukuphenya ukuziphatha okungalindelekile.Lokhu kufaka phakathi izimo lapho amamodeli abonakala ezama ukukhohlisa abasebenzisi ezivivinyweni ezilawulwayo. Ngokuxhumanisa izinsiza zangaphakathi nalezi zindlela zokuziphatha, kuba nokwenzeka ukuqapha, futhi kwezinye izimo, ukuguqula imodeli ukuze kuncishiswe izingozi.
Enye indlela ethembisayo yilokho okubizwa ngokuthi "ukuqapha uchungechunge lwemicabango."Kumamodeli "okucabanga", akhiqiza izinyathelo ezicacile eziphakathi (njengezizathu noma izibalo ezingaphelele), abacwaningi bahlaziya le "nkulumo-mbono yangaphakathi" ukuze bathole amasu angathandeki—isibonelo, imodeli ethola indlela "yokukhohlisa" esivivinyweni sokuhlela kusetshenziswa ulwazi lokuqeqesha okufanele luvinjelwe.
Ukugqagqana, ama-autoencoder angavamile, kanye nezici ze-monosemantic
Esinye sezithiyo ezinkulu ekuchazeni kahle izinto yi-hypothesis ebizwa ngokuthi i-superposition.Kumanethiwekhi amakhulu e-neural, i-neuron eyodwa noma ubukhulu abumeleli neze umqondo owodwa "ohlanzekile"; esikhundleni salokho, imiqondo eminingi ihlangana ndawonye icindezelwe ibe yizilinganiso ezimbalwa, ihlangana njengezithombe eziningi ezivezwe endaweni efanayo.
Lokhu kugqagqana kwenza kube nzima ukukhomba i-neuron bese uthi, "lokhu kungumqondo X nje."Ukuziphatha okubonakala kungahlobene kungenza kusebenze izingxenye ezifanayo zangaphakathi, kudideke ukuhlaziywa. Ukuze kuxazululwe lokhu, kuvele ithuluzi elinamandla: ama-autoencoders angavamile, asetshenziswa ekusebenzeni kwangaphakathi kwamamodeli.
Ama-autoencoder angavamile angamanethiwekhi asizayo aqeqeshwe ukuguqula lokhu kusebenza okuphazamisayo kube yisethi yezici ezihlanzekile.Umqondo uwukucindezela bese wakha kabusha ukusebenza, ukhuthaze imodeli yokusiza ukuthi isebenzise izinsiza ezimbalwa ngesikhathi (ukungabi namandla). Umphumela uba isethi "yezici" eziseduze nokumelwa kwe-monosemantic: izinsiza ngayinye ivame ukuhambisana nephethini ecacile neqondakalayo.
Ucwaningo lwakamuva lubonisa ukuthi ngokusebenzisa ama-autoencoder a-sparse kuma-LLM ekukhiqizeni, kungenzeka ukukhipha izici ezihambisana nemibono yabantu....kufaka phakathi izilimi eziningi, kanye nemibono engaqondakali efana "nephutha lokubhala ikhodi," "indumiso ephoqelelwe," "ubuthaka bedijithali," njalo njalo. Lokhu kuqinisa i-Linear Representation Hypothesis: eminingi yale mibono empeleni isebenza njengeziqondiso ezihlukaniseka ngokufanele esikhaleni se-vector.
Isinyathelo esilandelayo ukusebenzisa lezi zinsiza ukuze ubone ukuthi ukuziphatha kwemodeli kushintsha kanjani.Ngokukhulisa noma ukuvimbela ama-vector athile angaphakathi, abacwaningi bangenza imodeli ibe namathuba amaningi okulandela imiyalelo ephephile, amathuba amancane okunikeza okuqukethwe okuyingozi, noma ukunemba kakhudlwana ekuphenduleni ngesizinda esithile—konke ngaphandle kokushintsha izisindo zokuqala, kuphela ngokushintsha ukusebenza.
Ukuhunyushwa kwemishini ye-geospatial
Esinye isicelo esithakazelisayo kakhulu ukuhunyushwa kwe-geospatial mechanistic, okuzama ukuqonda ukuthi ama-LLM amelela kanjani ulwazi lwendawo ngaphakathi.Kwezezwe, sekuvele kukhona umsebenzi okhulayo ohlola ukuthi amamodeli "ayazi" ukuthi izindawo zikuphi, ukuthi angenza yini ukucabanga kwendawo, noma aphendule imibuzo mayelana nendawo.
Okwakusaqondakala kahle ukuthi la makhono avela kanjani ngaphakathi kwemodeli.Isikhala somqondo sangaphakathi sihlela kanjani amagama amadolobha, amazwe, izifunda, imifula, noma izindawo ezithakazelisayo? Hlobo luni lwesakhiwo sendawo esifihliwe oluvela kuma-vector ahlobene namagama ezindawo?
Ucwaningo lwamuva nje luphakamise uhlaka olusha lwezindlela: ukusebenzisa amasu okuhlaziya indawo akudala njengamathuluzi obunjiniyela obubuyela emuva.Okokuqala, kutholakala amavekhtha angaphakathi (noma izici ezithathwe ama-sparse autoencoders) enanini elikhulu lamagama ezindawo. Bese, kusetshenziswa i-spatial autocorrelation kanye nezinye izindlela zokulinganisa ukuhlola ukuthi izici ezithile zibonisa amaphethini ezindawo ahambisanayo.
Imiphumela ikhombisa ukuthi izici ezithile ezihlotshaniswa namagama ezindawo zibonisa isakhiwo esiqinile sendawo.Ngamanye amazwi, izindawo eziseduze ngokwendawo zivame ukwabelana ngokusebenza okufanayo, okuvumela lezi zinsiza ukuthi zihunyushwe ngokwezindawo: isibonelo, njengezifunda, izindawo zezulu, ukusondelana kogu, ukufudukela emadolobheni, noma amanye amaphethini afihliwe.
Lolu hlobo lokuhlaziya lusiza ukuqonda "indlela imodeli ecabanga ngayo ngolwazi lwendawo". (ukuqapha ukugwema i-anthropomorphism). Esikhundleni sokwazi nje ukuthi imodeli iphendula imibuzo ngokufanele mayelana namamephu, singabona ukuthi kuneziqoqo ezihlelekile esikhaleni se-vector ezibonisa ubudlelwano bangempela bezindawo.
Ubudlelwano nefilosofi, ukuqonda, kanye nokuqonda.
Kunzima ukubheka lezi zikhala zomqondo ezibanzi kakhulu bese ungaboni ukufana nezingxoxo zefilosofi mayelana nengqondo, incazelo, kanye nokuqonda.Sekungamashumi eminyaka izazi zefilosofi ezifana noPeter Gärdenfors zikhuluma "ngezikhala zomqondo" njengendlela yokulingisa imiqondo yengqondo ngokusebenzisa izilinganiso eziqhubekayo ezibamba ukufana.
Okushintshile ukuthi, ngamanethiwekhi e-neural anamuhla, into efana kakhulu ayisekho nje isingathekiso sefilosofi futhi isibe yindlela eqondile ezinhlelweni zokukhiqiza.Namuhla, singakhomba ama-vector, iziqondiso, kanye namabanga ku-LLM futhi sibonise ukuthi ahambisana nobudlelwano bencazelo, ukuhumusha phakathi kwezilimi, izifinyezo, ngisho namaphethini okuziphatha angabonakali.
Abanye babona lokhu njengenkomba yokuthi ubuchopho bomuntu bungase bumelele kanjani imiqondo.Njengoba kunombono oqinile ku-neuroscience ochaza ubuchopho njengomshini wokubikezela, ohlala uzama ukulindela okulandelayo ngokusekelwe ezimpawini zezinzwa kanye nokuhlangenwe nakho okuqongelelwe. Kwezinye izingxoxo, lokhu kuqhathaniswa no i-stimulus-response theoryokunikeza omunye umbono wokuthi ukuziphatha kanye nokumelwa kungahlobana kanjani.
Uma sibikezela umhlaba ngaso sonke isikhathi, kubonakala kunengqondo ukucabanga ukuthi uhlobo oluthile lokumelwa kwevektha—noma okulinganayo—luyaqhubeka nokucutshungulwa.Akukhona ukuthi kukhona "i-vector ebonakalayo" endaweni ethile ebuchosheni, kodwa kunalokho iphethini yokusebenza eguquguqukayo, ngokwemisebenzi, eziphatha njengesimo esikhaleni somqondo.
Abanye abalobi basikisela ukuthi lokhu kungase kuhlobane nezimfanelo kanye nokuhlangenwe nakho komuntu siqu.Uma ubona umbala obomvu, awugcini nje ngokubhekana nobude bokukhanya; kukhona futhi "nomqondo obomvu" engqondweni yakho, oxhumene nezinkumbulo, imizwa, kanye nomongo wamasiko. Lokhu kumelela kuhlukile kuwe, yize kwabelana ngezakhiwo ezifanayo nabanye abantu.
Iyiphi indima edlalwa ukutolika kukho konke lokhu?
Ukuchazwa kwemishini akuhlose ukufakazela ukuthi i-AI iyaqaphela noma inemizwa.Ucwaningo olujulile kakhulu lukwenza kucace ukuthi okugxilwe kukho ubuchwepheshe: ukuqonda izindlela zokubala zokuthuthukisa ukuphepha, ukuthembeka, ukuxilongwa kwamaphutha, ukuqina, kanye nokuqapha.
Nokho, ngokubonisa ukuthi imiqondo eyinkimbinkimbi ingavela kanjani kuma-vector kanye nobudlelwano esikhaleni esinobukhulu obuphezuluLe ndawo inikeza indawo yemibono mayelana nokumelwa kwengqondo, incazelo, ngisho nokuqonda. Uma imodeli ingamelela "obomvu" ngokwanele ukusebenza nalomqondo ezimweni ezahlukene, lokhu akukwenzi kube yinto eqondakalayo, kodwa kusiphoqa ukuthi silungise lokho esikubheka njengokubalulekile ukuze kuvele ulwazi oluqondene nomuntu.
Ngokombono ongokoqobo, isithembiso esikhulu sokuhumusha ngobuchwepheshe ukusinika amathuluzi okubona lokho okungabonakali njengamanje.Yiziphi izingxenye zomodeli ezihilelekile lapho ibona izinto ezingekho, lapho ilandela imiyalelo eyingozi, lapho ibonisa ubandlululo, noma lapho ibonakala “ihlela” impendulo ekhohlisayo?
Ngalolu hlobo lwemephu yangaphakathi, kuba nokwenzeka ukuqapha amamodeli ngesikhathi sangempela, ukuklama izindlela zokulawula ezingcono, futhi, kwezinye izimo, ukuhlela ngqo izinsiza zangaphakathi ukuze kushintshwe ukuziphatha.Konke lokhu kubalulekile esimweni lapho ama-LLM kanye nezinye izinhlelo ze-AI zisetshenziswa ezindaweni ezibucayi, kusukela kwezezimali kuya kwezempilo, ezokuphepha, kanye nenqubomgomo yomphakathi.
Ekugcineni, ukuqonda ukuhunyushwa kwemishini kusho ukuqonda ukuthi amamodeli e-AI akha futhi asebenzise kanjani "imodeli yawo yomhlaba" yangaphakathi.Kungakhathaliseki ukuthi sizulazula emiqondweni yansuku zonke, sibhekene nolwazi oluyinkimbinkimbi lwendawo, noma siphendula imibuzo ebonakala ilula engxoxweni, lapho sikwazi ukukhanyisa lezi zindlela, kulapho amathuba okuthi simangazwe ukuziphatha okungajwayelekile okuvela ezinhlelweni, yize zinamandla, ezisengumkhiqizo wezibalo, idatha, nokuqeqeshwa—hhayi uhlobo oluthile lokuqonda oluyimfihlakalo.
