I-Gemma 3n: Ibhizinisi elisha le-Google lokuletha i-AI ethuthukisiwe kunoma iyiphi idivayisi

I-Google ithathe isinyathelo esibalulekile emhlabeni wobuhlakani bokwenziwa nge Ukwethulwa kwe-Gemma 3n, imodeli ye-AI yomthombo ovulekile eklanyelwe ukusebenza kumadivayisi anomkhawulo wensiza. Lesi siphakamiso, okuyinto Manje ingadawunilodeka futhi ifakwe kumaselula, amaphilisi namalaptop., kusho ukuthi Ukufika kwe-multimodal AI entendeni yesandla sakho, ngisho nakumadivayisi ane-RAM engu-2 GB kuphela futhi akukho uxhumano lwe-inthanethi. Ukubukeka kwayo kwenzeka ngemva kokwethulwa kwayo ngesikhathi I-Google I/O yokugcina, futhi ibambe ukunaka konjiniyela nabasebenzisi abafuna izixazululo ze-AI zasendaweni, eziyimfihlo, nezisebenzayo.

Le modeli entsha isuselwe kunhloso ye Yenza kube lula ukufinyelela kumathuluzi obuhlakani bokwenziwa athuthukile ngaphandle kokuthembela kumaseva wamafuNgakho, i-Google ihlukanisa ngokucacile i-Gemma 3n kwezinye izindlela ezifana ne-Gemini, egcina indlela evaliwe futhi igxile kakhulu ekusetshenzisweni okuningi. Odabeni lukaGemma, kugxilwe ekuthuthukisweni okuvulekile kanye nocwaningo kanye nokusetshenziswa komuntu siqu kwe-AI, okuvumela ukuthi ilandwe, ilungiswe, futhi ihlanganiswe nenqwaba yezinhlelo zokusebenza.

Amakhono e-Multimodal nokusebenza kahle okuvelele

I-Gemma 3n igqame ikakhulukazi ngokuba yi-multimodal, okungukuthi, ingahumusha futhi ikhiqize umbhalo, izithombe, umsindo nevidiyo ngqo kusuka kudivayisi, ngaphandle kokusebenzisa ifu. Amandla ayo ayinhloko ahlanganisa ukuqaphela inkulumo, ukuloba, ukuhumusha, nokuhlaziya okubonakalayo kwesikhathi sangempela, okuyenza ifaneleke kahle imisebenzi yezemfundo, abasizi bomuntu siqu, noma amasistimu okuhumusha.

I-architecture eyakhiwe phezu kwayo, ebizwa ngokuthi I-MatFormer, ivumela imodeli ukuthi ihlukaniswe ngezinguqulo ezincane ezihlanganiswe ngaphakathi kweyinhloko, njenge-matryoshka. Ngenxa yalesi sakhiwo, I-Gemma 3n ingaphatha kangcono izinsiza futhi ivumelane nemikhawulo yehadiwe lapho isebenza khona.. Ngaphezu kwalokho, ihlanganisa i- ubuchwepheshe Ukushumeka Ngesendlalelo Ngasinye (PLE), lokho kunciphisa ukusetshenziswa kwememori ngaphandle kokulahlekelwa ukusebenza, ngaleyo ndlela ivumela ukuthi isebenze ngisho nakumadivayisi anezincazelo ezinesizotha.

I-Gemma 3n inikezwa ngezinhlobo ezimbili eziyinhloko: I-E2B y I-E4B, ngamapharamitha asebenzayo ayizigidigidi ezi-2.000 nezingu-4.000 ngokulandelanayo. Kodwa-ke, ngenxa yokwakheka kwawo, womabili amamodeli angasebenza nezidingo zenkumbulo ezilingana namamodeli amancane kakhulu, okuvula umnyango we-AI ethuthukisiwe kumadivayisi endabuko asezingeni eliphansi nelimaphakathi.

Okwe ukucubungula isithombe nevidiyo, i-Gemma 3n isebenzisa isifaki khodi I-MobileNet-V5, elungiselelwe ukuthi isebenze kahle ngisho nakumadivayisi eselula anamandla aphansi, okukuvumela ukuthi usebenze ngevidiyo ku-60 fps kumamodeli akamuva. Esigabeni somsindo, ivumela ukulotshwa kwezwi nokuhumusha okusheshayo, konke endaweni.

Ubumfihlo, ukusebenza nokutholakala

Ukusebenza ungaxhunyiwe ku-inthanethi ngokuphelele kungenye yamandla amakhulu e-Gemma 3n, Iqinisekisa ukuthi yonke idatha ecutshungulwe yi-AI ihlala kudivayisi ngokwayo, ngaleyo ndlela iqinisa ubumfihlo bomsebenzisi uma iqhathaniswa nezinye izixazululo ezisekelwe emafini. Lesi sici siphinde sihumushele ekusebenzeni kahle kwamandla okukhulu kanye nokusetshenziswa kwedatha okuphansi, izici ezibalulekile kumadivayisi eselula kanye nezindawo ezinokuxhumeka okulinganiselwe.

Ngokuphathelene nokusebenza, I-Gemma 3n isekela izilimi eziyi-140 zokucubungula amagama nezilimi ezingama-35 kwimodi yayo ye-multimodal.Ibonise ukusebenza okuvelele ekuhlolweni kwebhentshimakhi okufana ne-LMArena, lapho imodeli ye-E4B idlula amaphuzu angu-1.300, ibe ngeyokuqala ngamapharamitha angaphansi kwezigidigidi eziyi-10.000 ukufinyelela kuleli zinga.

I-Gemma 3n isikhona kakade iyatholakala kumapulatifomu amaningi konjiniyela, njenge-Google AI Studio, Hugging Face, Kaggle, nangamathuluzi afana ne-Google AI Edge noma i-Ollama. Idizayini yabo evulekile nokuvumelana nezimo zokuhlanganisa kwenza kube lula ukudala izinhlelo zokusebenza ezintsha ezifanelana nezidingo ezithile, kusukela kumasistimu ezemfundo kuya kubasizi abahlakaniphile namathuluzi okuhumusha angaxhunyiwe ku-inthanethi.

Ukuqhathanisa nezinye izindlela kanye nezinzuzo ezingokoqobo

Ukufika kwe-Gemma 3n kuza kumongo wokuvela kwe-AI yeselula ne-edge, Ezinye iziphakamiso zifaka i-Apple Neural Engine, i-Samsung Gauss, namamodeli avela ku-Meta ne-Microsoft. Nokho, nakuba eziningi zalezi zixazululo zidinga uxhumano lweseva, zinikeza umbhalo olinganiselwe noma amakhono wesithombe, noma zingavulekele ukuthuthukiswa kwangaphandle, i-Gemma 3n. Kuzibophezele ku-multimodality yangempela, ukungabi khona kokuncika kunethiwekhi nokuvuleleka emphakathini..

Izinzuzo eziphawuleka kakhulu kubasebenzisi kungenzeka sebenzisa i-AI ethuthukisiwe ngaphandle kokulahlekelwa ukulawula ubumfihlo, ujabulele impendulo esheshayo futhi unciphise izindleko ezihlobene nokusetshenziswa kwedatha yeselula. Kwabakhiqizi nabathuthukisi, i-Gemma 3n Imele ithuba lokuletha izinhlelo zokusebenza ezihlakaniphile ku-spectrum ebanzi kakhulu yamadivayisi, ngaphandle kokuthembela ku-hardware yakamuva noma ukuthuthukiswa kwememori okubizayo..

Umfutho we-Gemma 3n uze wagqugquzela abanye abakhiqizi ukuthi bakhulise umthamo we-RAM kumadivayisi abo amasha, belindele ukuhlanganiswa okukhulu kwekusasa kwe-AI yendawo. Ngakho, i-Google izibeka endaweni efanele emjahweni wokuzuza Ubuhlakani bokwenziwa obunamandla, obusebenzayo, obuvulekile, futhi obufinyeleleka ngempela.