Meta's 'data2vec' shine mataki na gaba zuwa Hanyar Sadarwar Jijiya ɗaya don Mulkin Su duka

Ana kan tseren ne don ƙirƙirar hanyar sadarwa guda ɗaya wacce za ta iya aiwatar da nau'ikan bayanai da yawa, ra'ayi na ƙarin ilimin wucin gadi wanda baya nuna bambanci game da nau'ikan bayanai amma a maimakon haka yana iya murƙushe su duka cikin tsari iri ɗaya.

Nau'in nau'i-nau'i iri-iri, kamar yadda ake kiran waɗannan hanyoyin sadarwa na jijiyoyi, ana ganin yawan aiki wanda aka wuce da bayanai daban-daban, kamar hoto, rubutu, da sautin magana, ta hanyar algorithm iri ɗaya don samar da maki akan gwaje-gwaje daban-daban kamar su. gane hoto, fahimtar harshe na halitta ko gano magana.

Kuma waɗannan cibiyoyin sadarwa masu ban sha'awa suna tattara maki akan gwaje-gwajen ma'auni na AI. Nasarar ta baya-bayan nan ita ce abin da ake kira 'data2vec,' wanda masu bincike a sashin Meta na AI, iyayen Facebook, Instagram, da WhatsApp suka kirkira. 

Batun, kamar yadda masana kimiyyar Meta, Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, da Michael Auli suka rubuta, shine kusanci wani abu mai kama da ikon koyo na gabaɗaya wanda tunanin ɗan adam ya ƙunshi.

"Yayin da mutane suka bayyana suna koyo ta irin wannan hanya ba tare da la'akari da yadda suke samun bayanai ba - ko suna amfani da gani ko sauti, misali," marubutan sun rubuta. a cikin shafin yanar gizo, "a halin yanzu akwai manyan bambance-bambance a cikin hanyar" hanyoyin sadarwar jijiyoyi suna sarrafa nau'ikan bayanai daban-daban kamar hotuna, magana, rubutu, "da sauran hanyoyin."

"Babban ra'ayin wannan hanyar," in ji data2vec, "shine don ƙarin koyo gabaɗaya: AI yakamata ya iya koyon yin ayyuka daban-daban, gami da waɗanda gaba ɗaya ba a sani ba."

Shugaban Meta, Mark Zuckerberg, ya ba da tsokaci game da aikin, yana ɗaure shi zuwa Metaverse na gaba:

Ci gaba mai ban sha'awa: Binciken Meta AI ya gina tsarin da ke koyo daga magana, hangen nesa da rubutu ba tare da buƙatar bayanan horo mai lakabi ba. Mutane sun fuskanci duniya ta hanyar haɗakar gani, sauti da kalmomi, kuma tsarin irin wannan na iya wata rana fahimtar duniya yadda muke yi. Wannan duk daga ƙarshe za a gina shi cikin gilashin AR tare da mataimaki na AI don haka, alal misali, zai iya taimaka muku dafa abincin dare, lura idan kun rasa wani sashi, yana sa ku kashe zafi, ko ƙarin ayyuka masu rikitarwa.

Sunan data2vec wasa ne akan sunan shirin don "haɓaka harshe" An inganta shi a Google a cikin 2013 ake kira "word2vec." Wannan shirin ya annabta yadda kalmomi ke taruwa tare, don haka word2vec wakilci ne na hanyar sadarwa na jijiyoyi da aka tsara don takamaiman nau'in bayanai, a cikin wannan yanayin rubutu. 

Hakanan: Bude ƙofofin pod bay, don Allah, HAL: Meta's AI yana kwatanta karatun leɓe

Dangane da data2vec, duk da haka, Baevski da abokan aikinsa suna ɗaukar daidaitaccen sigar abin da ake kira Transformer, wanda Ashish Vaswani da abokan aikinsa suka haɓaka. Google a cikin 2017 da kuma fadada shi don amfani da shi don nau'ikan bayanai da yawa. 

Cibiyar sadarwa ta Transformer ta asali an ƙirƙira ta ne don ayyukan harshe, amma an daidaita ta sosai a cikin shekarun da suka gabata don nau'ikan bayanai da yawa. Baevski et al. nuna cewa za a iya amfani da Transformer don aiwatar da nau'ikan bayanai da yawa ba tare da an canza su ba, kuma cibiyar sadarwar da aka horar da ita wacce ke haifar da iya yin ayyuka daban-daban. 

A cikin takarda na yau da kullun, "data2vec: Babban Tsari don Koyon Kula da Kai a cikin Magana, hangen nesa da Harshe, "Baevski et al., horar da Transformer don bayanan hoto, sautin murya na magana, da wakilcin harshen rubutu. 

Data2vec shine "algorithm na farko wanda ke kula da kansa wanda ke aiki don hanyoyi masu yawa, wato magana, hangen nesa, da rubutu," rubuta Baevski da tawagar a cikin shafin yanar gizon.

Babban Transformer ya zama abin da ake kira pre-training wanda za a iya amfani da shi a kan takamaiman hanyoyin sadarwa na jijiyoyi don aiwatar da takamaiman ayyuka. Misali, marubutan suna amfani da data2vec azaman horo na farko don samar da abin da ake kira “ViT,” “Mai canza hangen nesa,” hanyar sadarwa ta jijiyoyi da aka tsara musamman don ayyukan hangen nesa waɗanda an gabatar da shi a bara by Alexey Dosovitskiy da abokan aiki a Google. 

Lokacin da aka yi amfani da shi akan ViT don ƙoƙarin warware madaidaicin gwajin ImageNet na tantance hoto, sakamakon su yana zuwa a saman fakitin, tare da daidaiton 84.1%, fiye da maki na 83.2% da wata ƙungiya a Microsoft ta samu waɗanda suka riga sun horar da su. ViT, wanda Hangbo Bao ya jagoranta, bara.

Kuma data2vec Transformer iri ɗaya yana fitar da sakamako waɗanda ke da yanayin fasaha don fahimtar magana kuma masu gasa, idan ba mafi kyau ba, don koyan harshe na halitta:

Sakamakon gwaji ya nuna data2vec don zama mai tasiri a cikin dukkanin hanyoyi guda uku, saita sabon yanayin fasaha don ViT-B da ViT-L akan ImageNet-1K, ingantawa a kan mafi kyawun aikin da ya gabata a cikin sarrafa magana akan fahimtar magana da yin daidai da RoBERTa. akan ma'auni na fahimtar harshen GLUE. 

Babban abin da ke faruwa shi ne cewa wannan yana faruwa ba tare da wani gyare-gyare na hanyar sadarwa na jijiyoyi don zama game da hotuna ba, kuma iri ɗaya don magana da rubutu. Madadin haka, kowane nau'in shigarwa yana shiga cikin hanyar sadarwa iri ɗaya, kuma yana kammala aiki na gaba ɗaya. Wannan aikin shine aikin da hanyoyin sadarwa na Transformer ke amfani da su koyaushe, wanda aka sani da "masked Hasashen." 

Hakanan: Supermodel na Google: DeepMind Perceiver mataki ne akan hanyar zuwa injin AI wanda zai iya sarrafa komai da komai.

Yadda data2vec ke aiwatar da tsinkayar abin rufe fuska, duk da haka, hanya ce da aka sani da koyo "mai kulawa da kai". A cikin tsarin kulawa da kai, ana horar da hanyar sadarwa ta jijiyoyi, ko haɓaka, ta hanyar wucewa ta matakai da yawa. 

Na farko, hanyar sadarwar tana gina wakilcin yuwuwar haɗin gwiwa na shigar da bayanai, zama hotuna ko magana ko rubutu. Bayan haka, sigar hanyar sadarwa ta biyu tana da wasu abubuwan shigar da bayanan “wanda aka rufe,” ba a bayyana su ba. Dole ne ya sake gina yuwuwar haɗin gwiwa wanda sigar farko ta hanyar sadarwar ta gina, wanda ke tilasta masa ƙirƙirar mafi kyawun kuma mafi kyawun wakilcin bayanan ta ainihin cika guraben. 

Cibiyoyin sadarwa guda biyu, wanda ke da cikakken tsari na yuwuwar haɗin gwiwa, da kuma wanda ke da sigar da ba ta cika ba wanda yake ƙoƙarin kammalawa, ana kiranta da hankali sosai, “Malami” da “Student.” Cibiyar sadarwar ɗalibi tana ƙoƙarin haɓaka ma'anarta na bayanai, idan kuna so, ta hanyar sake gina abin da Malamin ya rigaya ya cimma.

Za ka iya duba lambar don samfuran akan Github.

Ta yaya neural network ke yin malami da ɗalibi don nau'ikan bayanai daban-daban guda uku? Makullin shine "manufa" na yiwuwar haɗin gwiwa, a cikin dukkanin bayanai guda uku, ba takamaiman nau'in bayanan fitarwa ba ne, kamar yadda yake a cikin sigogin Transformer don takamaiman nau'in bayanai, kamar Google's BERT ko OpenAI's GPT-3 . 

Madadin haka, data2vec yana ɗaukar wasu ɗimbin yadudduka na hanyoyin sadarwa waɗanda suke ciki cibiyar sadarwar jijiyoyi, wani wuri a tsakiya, wanda ke wakiltar bayanai kafin a samar da shi a matsayin fitarwa ta ƙarshe. 

Kamar yadda marubutan suka rubuta, “Daya daga cikin manyan bambance-bambancen hanyarmu […] ban da aiwatar da hasashen abin rufe fuska, shine amfani da maƙasudai waɗanda suka dogara akan matsakaicin yadudduka da yawa daga cibiyar sadarwar malamai.” Musamman, "muna mayar da wakilcin Layer cibiyar sadarwa na jijiyoyi maimakon kawai saman Layer," don haka "data2vec yana tsinkayar bayanan bayanan shigarwa."

Sun kara da cewa, "Muna amfani da fitarwa na FFN [cibiyar ciyarwa ta gaba] kafin haɗi na ƙarshe a cikin kowane toshe a matsayin manufa," inda "toshe" shine Transformer daidai da Layer na cibiyar sadarwa.

Abin nufi shi ne, duk nau’in bayanan da ke shiga ya zama kalubale iri daya ga cibiyar sadarwa ta dalibai na sake gina wani abu a cikin cibiyar sadarwa ta jijiyar da Malamin ya hada.

Wannan matsakaicin ya bambanta da sauran hanyoyin kwanan nan don gina hanyar sadarwa ɗaya don Crunch Duk Bayanai. Misali, a bazarar da ta gabata, rukunin DeepMind na Google ya ba da abin da ya kira “Perceiver,” nau'in nasa nau'ikan nau'ikan nau'ikan Transformer. Horar da hanyar sadarwa na Perceiver neural shine mafi daidaitaccen tsari na samar da fitarwa wanda shine amsar mai lakabi, aikin kulawa kamar ImageNet. A cikin tsarin kulawa da kai, data2vec baya amfani da waɗancan alamun, yana ƙoƙarin sake gina bayanan cikin gida na cibiyar sadarwa. 

Ko da ƙarin ƙoƙarce-ƙoƙarce mai ban sha'awa yana cikin fuka-fuki. Jeff Dean, shugaban kokarin Google na AI, a watan Oktoba ya yi ba'a game da "Hanyoyin," abin da Dean ya yi iƙirari shine "na gaba tsara AI architecture"don sarrafa bayanai masu yawa.

Yi la'akari, tsarin data2vec na gaba ɗaya zuwa hanyar hanyar sadarwa guda ɗaya don hanyoyi masu yawa har yanzu yana da bayanai da yawa game da nau'ikan bayanai daban-daban. Hoto, magana da rubutu duk an shirya su ta hanyar aiwatar da bayanan. Ta wannan hanyar, yanayin tsarin hanyar sadarwa da yawa har yanzu yana dogara ne akan alamu game da bayanan, abin da ƙungiyar ke nufi da "ƙananan maƙallan shigar da takamaiman tsari."

Hakanan: Google ya buɗe 'Hanyoyin Hannu', AI mai zuwa na gaba wanda za'a iya horar da shi zuwa ayyuka da yawa

"Duk da tsarin ilmantarwa bai ɗaya, har yanzu muna amfani da ƙayyadaddun kayan aikin cirewa da dabarun rufe fuska," in ji su.

Don haka, har yanzu ba mu kai ga duniyar da aka horar da net ɗin jijiyoyi ba tare da ma'ana kowane nau'in bayanan shigarwa ba. Har ila yau, ba mu kasance a lokacin da cibiyar sadarwa ta jijiyoyi za ta iya gina wakilci ɗaya wanda ya haɗu da duk nau'ikan bayanai daban-daban ba, ta yadda jijiyoyi ke koyan abubuwa a hade.

An bayyana wannan gaskiyar daga musayar tsakanin ZDNet da marubuta. ZDNet sun isa ga Baevski da tawagar kuma suka tambaya, "Shin wakilcin latent da ke aiki a matsayin maƙasudi an haɗa su tare da duk hanyoyin guda uku a kowane mataki na lokaci, ko yawanci ɗaya ne daga cikin hanyoyin?"

Baevski da tawagar amsa cewa shi ne na karshen hali, da kuma su reply yana da ban sha'awa a faɗi a tsayi:

Matsalolin latent ba haɗe-haɗe ba ne don hanyoyin uku. Muna horar da samfura daban-daban don kowane tsari amma tsarin da samfuran ke koya iri ɗaya ne. Wannan shi ne babban sabon aikinmu tun da a baya an sami babban bambance-bambance a yadda ake horar da samfura ta hanyoyi daban-daban. Masana kimiyyar jijiyoyi kuma sun yi imanin cewa mutane suna koyo ta hanyoyi iri ɗaya game da sautuna da duniyar gani. Ayyukanmu yana nuna cewa koyo mai kulawa zai iya aiki iri ɗaya don hanyoyi daban-daban.

Ganin ƙayyadaddun ƙayyadaddun ƙayyadaddun yanayin data2vec, hanyar sadarwar jijiya wacce zata iya kasancewa da gaske Networkaya Daya Don Mallakar Su Duka ya kasance fasahar nan gaba.