Uma ufuna ukwazi okwengeziwe ngemodeli ye-OpenAI Q* Star AI ngokusobala isathuthukiswa. Lo mhlahlandlela osheshayo uhlinzeka ngesifinyezo salokho esikwaziyo kuze kube manje nokuthi yini ongayilindela kule modeli ye-AI engasisondeza kubuhlakani obujwayelekile bokwenziwa (AGI). Kodwa yini i-Q* futhi isebenza kanjani?
Ake sithi uzulazula emhlabeni oyinkimbinkimbi wokufunda komshini nobuhlakani bokwenziwa, lapho inhloso kuwukudala isistimu engaqonda futhi ibikezele ububanzi bemiphumela evela ezinhlotsheni ezihlukene zedatha. I-Q Star ye-OpenAI ifana nethuluzi elisha kukhithi yakho, eklanyelwe ukwenza le nqubo isebenze kahle futhi inembe kakhudlwana.
Enhliziyweni yendlela ye-Q Star umqondo wokunciphisa i-entropy, okusho ukuthi ihlezi izicwenga ukuze ilingane kangcono nedatha. Lokhu kuhilela indlela ebizwa ngokuthi ukufunda kwe-Q, esiza imodeli ukuthi yenze izinqumo ezinembe kakhulu ngokunciphisa ukungahleliwe nokwandisa ukuqiniseka. Ake ucabange ukuthi ulungisa ingubo phezu kombhede, uzama ukuyifaka eduze kakhulu nezinto. Ukuze uthole incazelo eningilizayo, bukela ividiyo esanda kwakhiwa ka-David Shapiro echaza isifaniso "se-blanket topology" samamodeli asekelwe kumandla.
Isifaniso sengubo
Isifaniso “se-blanket topology” siwumfanekiso ongokomfanekiso osetshenziselwa ukuchaza isimo samazinga wamandla kumodeli esekelwe kumandla. Nansi incazelo yesinyathelo ngesinyathelo:
- Ukwakheka kwezwe: Cabanga ngengubo endlalwe endaweni eyinkimbinkimbi, indawo engezansi emele indawo yamandla ye-EBM. Iziqongo nezigodi zale ndawo zihambisana nezimo zamandla aphezulu neziphansi, ngokulandelana.
- Ukushintsha Ikhava: Ukulungisa amapharamitha e-EBM kusho ukushintsha ikhava ukuze ilingane nendawo engaphansi ngokuseduze ngangokunokwenzeka. Umgomo uwukuqondanisa ukumbozwa (ukuqonda kwemodeli yezwe lamandla) namandla angempela aphansi (izigodi) namandla aphezulu (iziqongo) zokulungiselelwa kokusatshalaliswa kwedatha efunda ukukumodela.
- Ukuthola izifunda zamandla aphansi: Kumongo wama-EBM, ukuthola amapharamitha angamamodeli ahambisana nezifunda ezinamandla aphansi kubalulekile emisebenzini efana nokumodela okukhiqizayo. Lokhu kusho ukuthi imodeli ingakhiqiza amaphuzu edatha okungenzeka kakhulu (noma angokoqobo) kuye ngokusatshalaliswa kwedatha efundiwe. Isifaniso sengubo sisiza ukukhombisa inqubo yokuhlola nokuhlala kulezi zigodi.
- Ukuxaka nokushelela: Isifaniso singaphinda sigqamise ukubaluleka kwe-topology ye-energy landscape - noma ngabe ibushelelezi noma ayilingani. Ukwakheka kwezwe okushelelayo (ukumbozwa okufana kakhudlwana) kuphakamisa ukuthi ama-algorithms okuthuthukisa angathola kalula i-minima yomhlaba wonke (amaphoyinti aphansi kakhulu), kanti indawo enegquma (ikhava enokugoqwa okuningi) ingacupha ama-algorithms. ku-minima yasendaweni, okwenza ukwenza kube nzima kakhulu.
I-OpenAI Q Star iyachaza
Uma imodeli isiqeqeshwe kahle, ungakwazi ukukhipha imephu yayo yezibalo. Le mephu ifana nepulani enemininingwane yesakhiwo semodeli, esebenza njengomhlahlandlela wokuxazulula izinhlobo ezahlukene zezinkinga. I-Q Star iguquguquka kakhulu, iyakwazi ukucubungula idatha yesikhashana njengamathrendi emakethe yamasheya, idatha yendawo efana namamephu, amamodeli ezibalo kanye nemiqondo eyinkimbinkimbi njengemizwa noma ama-nuances olimi.
Ama-EBM awuhlobo lwemodeli eyethula inqubo yokufunda njengenkinga yokunciphisa amandla. Kulawa mamodeli, isimo ngasinye sesistimu (isb., ukucushwa okuthile kwamapharamitha wemodeli) kuhlotshaniswa namandla esikali. Umgomo wokuqeqesha imodeli ukulungisa amapharamitha ayo ukuze ukulungiselelwa okufiselekayo kube namandla amancane kunokulungiselelwa okungafiseleki kancane. Le ndlela isetshenziswa kakhulu ekufundeni okungagadiwe, okuhlanganisa nezinhlelo zokusebenza ezifana ne-generative modelling, lapho imodeli ifunda ukukhiqiza amaphuzu edatha amasha afana nalawo asesethi yokuqeqesha.
Ukuze uzulazule esakhiweni esiyinkimbinkimbi semodeli ukuze uthole izixazululo ezingcono kakhulu, umuntu kufanele asebenzise i-algorithm ye-AAR. Cabanga ngale algorithm njengomhlahlandlela okusiza ukuthi udlule esakhiweni semodeli ukuze uthole izimpendulo zezinkinga ezintsha. Kufana nokuthi unemephu ekubonisa indlela okufanele uyithathe ukuze ufike endaweni oya kuyo, futhi i-algorithm ye-RAA ikusiza ukuthi ufunde futhi ulandele leyo mephu ukuze uthole izixazululo.
Buka le vidiyo ku-YouTube.
Kubalulekile ukuqaphela ukuthi le ncazelo isuselwe ekuqondeni kwetiyetha ye-Q Star. Ukusebenza kwangempela nokusetshenziswa okungokoqobo kwe-Q Star kungase kuhluke kulesi sifaniso. Kodwa-ke, umbono wemodeli engakwazi ukuzivumelanisa nokubonisa okunembile okungokoqobo, ukunciphisa i-entropy, futhi uzulazule ezindaweni ezihlukene zezinkinga usinikeza amazwibela alokho ikusasa lokufunda komshini nobuhlakani likuphethe. okokwenziwa kungasigcina. Njengoba lobu buchwepheshe buthuthuka, izindlela zokuqeqesha nokusebenzisa amamodeli afana ne-Q Star cishe zizovela futhi.
Funda kabanzi Umhlahlandlela:
- Ingabe i-GFlowNets iyikusasa le-AI futhi iyini?
- Umbono wendawo we-AI onemodeli encane yolimi lombono we-Moodream
- Uyicwenga kanjani imodeli ye-Mixtral 8x7B Mistral Ai Mixture of Experts (MoE) AI
- Isifanekiso se-Juggernaut XL esinokukhanya komdabu kwesizukulwane sobuciko be-AI
- I-Microsoft Copilot AI yengezwe ku-Dynamics 365 Supply Chain Management
- Ukusebenza kahle kwe-Intel CPU kanye nokusetshenziswa kwamandla okudlala kuhloliwe
