Ukuhlolwa kwe-MIT Ukubonisa ukuthi i-AI ingathola imodeli ezingaphezu kwe-AI

Yini inguqulo enkulu phakathi kwe-AI model ne-brain yabantu?

Ngesikhathi eside, izixazululo ezinguquguqukayo ziye ziye zithunyelwe – umzimba uyenziwe kakhulu emoyeni, i-multi-faceted in its media of input, futhi i-chemically enikezelwe ngaphandle kokusebenza kwamakhemikhali – kodwa isakhiwo esikhulu se-brain ye-human kuyinto plasticity yayo enhle. Uma isakhiwo se-body ye-patient (njenge-fingers, isikhunta, noma izindandatho eziningi) iyahlukaniswa, isakhiwo se-sensorimotor ye-neural esekelwe kwelinye isakhiwo se-body, manje ngaphandle kwe-nerve ekupheleni ukuxhumana, kuyaqala ngokushesha, ne-neurons "ukushintsha" ukuze asize nezinye izakhiwo ze-nerve ekulawula.WazeI-Plasticity inikeza abantu ukufundisa izinzuzo nezobuchwepheshe: njengoba kubizwa, "i-neurons that fire together wire together". I-muscle memory kanye ne-near-instant fact-recall zihlanganisa izindawo ezimbili ze-plasticity-enabled of our lives that we could never live without. For decades, abacwaningi abakwazi ukufundisa umsebenzi elifanayo kumamodeli we-AI - kuze kube manje. On June 12, a team of MIT abacwaningi wabhala iphepha yokuhlola okuhlobisa ukuthi inqubo ye-AI kungase ngokuvamile usebenzisa izinhlelo zokufundisa afana ne-human-like ukuzeUkuphucula ukusebenza wakhoKule nqakraza, sincoma imiphumela yemoral ne-technological ye-so-called Self-Adapting Language Model (SEAL), i-AI yokuqala ehlabathini.

Ukufundwa okungagunyaziwe

Ngokuvamile, amamodeli we-AI abasebenzisa isakhiwo se-Transformer angakwazi ukufundisa izicelo ezithile, kodwa izindlela eziningana ezinikezele ezingenalutho futhi engaphansi kokusebenza kakhulu. Mhlawumbe indlela engcono kakhulu ukufundisa imodeli ukuba usebenza izinzuzo ezithile - njenge-translate english ku-Chinese noma ukwenza imibuzo ye-trigonometry ngokucacileyo - iye isebenzisa inqubo ebizwa ngokuthi i-Supervised Fine Tuning, noma i-SFT ngokucacisa. Lolu hlobo ilungiselele ngezansi:

• Ukuhlola umsebenzi esifanele ufuna ukwenza SFT ku. Njengoba isibonelo, sicela ukuthatha isibonelo yokwenza umbhalo umbhalo omtsha.
• Ukuze isibonelo lethu, indlela enhle kodwa enhle ukwenza lokhu kuyinto nje ukusetshenziswa umbhalo singamafutha eyenziwe e-internet futhi ukuxhaswa ngama-resumers enhle we-content ne-characteristics of the songs.
• Ukusebenza kwe-SFT ku-model. Lokhu kusebenza ngokuvamile nge-process ebizwa ngokuthi i-Gradient Descent, ingxenye yobuchwepheshe ebonakalayo kulesi sihloko. Phakathi nenani elikhulu le-training iterations, le nkqubo ivimbele ama-weights ye-model ngokuvamile ukuze inokukwazi ukukhiqiza into efana ne-output (i-lyrics ye-song) ngokuvumelana ne-input yayo (ukubhali oluthile le-song).

Ukuze zonke izidingo zayo kanye nezidingo zayo, i-SFT yenza umsebenzi, okungenani isixhobo ngaphakathi kwe-repertoire ye-AI developer yokufaka izixazululo zokhuseleko ezithile noma ukuphucula ukusebenza kwe-AI ku-tasks ezithile. Ngokuqhathanisa, ngokwemvelo se-SFT yenza inqubo enhle futhi enhle, okwenziwe ngokuvamile inani elikhulu le-high-quality data ebonakalayo endaweni yokuphendula (isib. I-Mathematical reasoning, i-Grammatical style). Nakuba izifundo ezininzi zophando zibonise ukuthi i-SFT ezivamile ingasetshenziswa ngokufanelekileyo, idatha eyenziwe nge-AI, i-SFT ikhona isixhobo esebenzayo ngokuzimela, njengoba ukugu

Izinzuzo ze-Evolution

Qaphela: Lezi zibonakalayo kulesiqingatha kubaluleke esithathwe kusuka ku-June MIT study "I-Self-Adapting Large Language Models" by Zweiger et al.

One of the disadvantages of traditional SFT has always been the human effort involved—SFT pipelines often had to be handcrafted by human AI researchers, ngisho nangokuthi iyindlela efanelekayo yokuguqulwa imodeli elilodwa ukuyenza ngokunemba okungenani izinhlobo zokusebenza. Ukusabela ngempumelelo eminyakeni omusha ku-synthetic data, abacwaningi wahlukanisa umqondo yokusebenzisa kuphela i-AI-generated SFT idatha, ngokushesha ukhangela umbuzo ukuthi abantu kungenziwa ngaphandle kwe-SFT loop ngokugcwele. Ukusabela kwabo, i-Self-Adapting Language Model (SEAL), kuyinto, ngokuvamile, ingxenye ye-framework engaphezu kuka-pre-trained SFT dataI-decoder-only i-transformer model(Ukuhlola usebenzisa amamodeli amabili amaphrojekthi, i-LLaMa-3.2 ne-Qwen-2.5B, izimo zokuhlola ezahlukile), i-software ye-”tool execution” kanye ne-SEAL nethishini ngokuvamile, ngokuvumelana nenkinga lokuphendula kwezinye imibuzo ye-benchmarking (i-SEAL).IsikhokeloI-SEAL inethiwekhi akuyona ngokuvamile futhi akhiqize inkinobho lokuphendula – kunzima ukwenza i-SFT ku-decoder-only transformer model kanye nokukhuthaza ukucindezeleka kwe-transformer.ImodeliUkwenza lokhu, inethiwekhi ye-SEAL iboniswe izixhobo ezimbili eziphambili:

• I-Synthetic Data Generation: Ngokusebenzisa le tool, inethiwekhi elinye izihlanganisa isixhumanisi (ngokuqondene ne-prompt) nokukhiqiza ama-SFT ama-pairs. Ngokwesibonelo, uma ukunikezelwa kwegama lokuthuthukiswa kwegama le-airplane, i-tuning pair ingaba ("Yini indiza ye-jet yokuthengisa yokuqala?" "i-De Havilland Comet"). Nakuba ifomu le-question-and-answer isetshenziswe ngokuvamile, le tool ingathuthukisa izinhlobo ezahlukahlukene zensimbi ukuze kuhlobene nezidingo zokusebenza ezithile.
• I-Hyperparameter tuning: Njengamanje, i-SFT kuyinto inqubo enikezayo eziningana nezinyathelo eziningana; Ngakho-ke izakhiwo eziningana nezinyathelo zokusebenza zitholakala ku-process eyenziwa ngokuthi i-hyperparameter tuning. Ngokuba ushiye le tool, i-SEAL ingakwazi ukuqala i-SFT nge izakhiwo ezithile (njenge-Learning Rate, # ye-Epochs (i-iterations), noma i-batch side ye-Gradient Descent), ngokuvamile ukuguqulwa ukuthi i-decoder iyahlekile (noma enhle).

Ngemva ukuthi i-SEAL has izindlela ezimbili ezinzima ezinikezela ukufundisa imodeli ye-AI, kufuneka kuphela ukuqeqeshwa indlela yokusebenzisa. Ekubeni ukuqeqeshwa kwayo, i-SEAL isetshenziselwa izixhobo ezimbili ngempumelelo kuzo zonke izicelo ze-benchmarking ezivela ku-framework. Lezi zokusebenza okuzenzakalelayo (i-SEs, njengoba abacwaningi wabhala) ngeke ikhiqize idatha ye-contextual, kodwa ayikho i-verbatim, ye-fine-tuning ngaphakathi kwegama le-prompt kanye nokuguqulwa imodeli yokuqala ye-decoder-only ngokusebenzisa izinyathelo ze-hyperparameter tuning, okwenza inethiwekhi ukukhiqiza imiphumela eyahlukile kunazo ngaphambi. Nokho, kukhona isUkuhlobisaUkwakhiwa kwe-benchmarking iyatholakala ku-"internal loop", eyenziwe nge-model θ entsha kanye ne-benchmarking yokuqala. Uma imodeli, ekuphenduleni kwebhizinisi le-benchmarking,NgaphezuNgaphezu kokubili kwimodeli ye-benchmarking, i-"internal loop" inikeza isignali ye-reward ye-positive. Uma izincazelo zihlukile, akukwazi ukuguqulwa kokubili; uma i- θ' iboniswa ukuthi iyahlekile ngokusekelwe ku-benchmarking question, ivumela ukuguqulwa kwe-negative. Ngokuye, le nkqubo ikakhulukazi ngokuvamile nge-example ye-Reinforcement Learning, lapho i-SEs ezinhle "ukuguqulwa" nge-reward ye-positive futhi i-SEs ezinhle zihlukile ngokufanayo; ngokusebenzisa ama-iterations ezininzi le-training, i-SEAL iboniswa kahle ekusebenziseni i-decoder ngokusebenzisa ama-self-edits. Enye

Ukwakhiwa kwezinkomba ezintsha zokusebenza ezinzima, ikakhulukazi ngenxa yokuzonwabisa kakhulu ukuze ukuqinisekisa ukuthi ukufundisa akuyimfuneko engabonakali ngokufanelekileyo noma imiphumela emibi "ukushintshwa" phakathi kwamakhemikhali. Abacwaningi abacwaningi abacwaningi zihlanganisa lezi zimpumelelo ngokusebenzisa ama-decoder-only transformer amamodeliNgaphandleNgaphezu kwalokho, imodeli ivimbele ukuthi izivivinyo ze-benchmarking asetshenziselwa, okungenani ukuthi izivivinyo ze-training zindiza ezidlulile zihlanganisa zonke izimo, ngokuvamile ukunciphisa amathuba ukuthi imodeli kuphela "ukufunda isivivinyo". Ngaphezu kwalokho, imodeli ivimbele ukuthi izivivinyo ze- θ' zihlanganisa ngokuphelele nangokuthi ku- θ futhi ukuthi imodeli yokuqala ayibhalwe ngezikhathi ezivamile, ivimbele ukuthi zonke izivivinyo ze-SEAL zihlanganisa i-SFT ukuze zihlanganise isibonelo entsha se- θ', kuyatholakala ku- θ esifanayo.

Iziphumo ziye zihlanganisa; Enye test benchmarking eyenzelwe yi-researchers, imodeli ivela izinga lokuphumula we-72.5%, ukusuka ku-0% ngaphandle kwe-SEAL fin-tuning, okuvumela umthamo enhle ye-framework yayo. Uma i-refined ne-integrated ngokugcwele, le framework ingangena isakhiwo esitsha ekuphuculeni ukusebenza kwe-AI emakethe ezithile noma ngokuvamile.

Umehluko lithunyelwe kwami ngu-Our AI, isakhiwo se-AI Ethics esekelwe abafundi kanye nesifundo se-student eyenza ukuhlinzeka iziphakamiso ze-AI ngaphandle kwe-inthanethi ezivamehluko ezintsha. Uma ungenza le nqaku, sicela uchofoze i-publikations yethu yenyanga kanye nama-articles ezizodwa ku-https://www.our-ai.org/ai-nexus/read!

Umehluko lithunyelwe kwami ngu-Our AI, isakhiwo se-AI Ethics esekelwe abafundi kanye nesifundo se-student eyenza ukuhlinzeka iziphakamiso ze-AI ngaphandle kwe-inthanethi ezivamehluko ezintsha. Uma ungenza le nqaku, sicela uchofoze i-publikations yethu yenyanga kanye nama-articles ezizodwa ku-https://www.our-ai.org/ai-nexus/read!

Ukulungiselela, noma awukwazi ukufunda?

Ngaphandle kokuphumelela kwe-technically, imiphumela ye-team ye-research iyatholakala, iziphumo ezingu-sociological kanye ne-philosophical ezingenalutho okuhlolwa kwalo akuyona akuyona. Ngingathanda ngokuqhubekayo izivakashi ze-biological computing (bheka:UkuhlobisaukusukaMay edition of I-AI Nexus Magazine) ngoba ngicabanga ukuthi ama-cluster ye-neuronal, njengezinto ezisetshenziswe ku-computer ye-biological, zihlanganisa nezinsizakalo zendalo ngenxa yokuba zihlanganisa umthamo we-consciousness, futhi, ngisho nangokungabikho, zihlanganisa ukuba zihlanganisa ngokwemvelo ngenxa ye-plasticity. I-SEAL kuyinto ebalulekile ngaphandle kwe-methode yokuphucula ukusebenza kwe-model ku-benchmark tasks; kuyinto inkqubo yokulungisa yokuqala yokufakelwa kwe-AI lapho imodeli ye-AI ibonise ngempumelelo umthamo wokulungisa ngqo i-AI elinye. Akukho kuphela lokhu kubonisa ukuthi singathanda emzimbeni esizayo e-AI eyenza indawo ye-AGI singularity, kunikeza imibuzo yomsindo

Kuyinto ukwahlukanisa kufakwe nge-adaptability kanye ne-consciousness. Thina siphinde lula ukufinyelela ku-blade of grass njengoba thina ufunde ukuthi, nangona kungabangela ukuphazamiseka, akuyona inkulumo wama-animalistic ye-pine njengoba ayikho ama-nerves. Nokho, i-blade of grassWazeNgaphezu kwalokho, sinikeza ukuchithwa kwama-animal, futhi bakwazi ukucubungula ukucubungula emkhakheni yayo ngokuvamile ngokuvimbela ngokuvimbela kwama-plate ye-concrete. Kodwa-ke, sinikeza ukuchithwa kwama-animal, futhi ngitholile ukuthi lokhu kungenzeka ngoba sinikezele ukuthi ukucubungula ukucubungula inikeza ukuguqulwa okungenani okungenani kakhulu – ukuchithwa noma ukuchithwa, ngisho – okuyinto abantu, okuyinto ama-animal ngokuvimbela okufanayo ku-ukudluliselwa ku-ukudluliselwa, zihlanganisa. Izilwane zihlanganisa ukucubungula-ukudluliselwa-ukudluliselwa-ukudluliselwa-ukudlUkulungiselela ngempumelelo umlilo wabantuukuthi i-humane eminingi angaphezu kuka-70% yamafutha.

Uma imodeli ye-AI isebenza njenge-umuntu kuzo zonke izindawo, ungatholakala njengomphakathi? Ingaba ukucubungula kwe-AI uzodinga amamodeli ezizodwa nezimo ezingenalutho ukuthi ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye “umzimba”? I-time kuphela ingathola.

Umbhali: Thomas Yin