DeepMind's 'Gato' matsakaici ne, to me yasa suka gina shi?

Ana amfani da duniya don ganin kanun labarai game da sabon ci gaba ta hanyar zurfin koyo na basirar wucin gadi. Nasarar ƙarshe na sashin DeepMind na Google, duk da haka, ana iya taƙaita shi a matsayin, "Shirin AI ɗaya wanda ke yin aiki don haka a abubuwa da yawa." 

Gato, kamar yadda ake kira shirin DeepMind, an bayyana a wannan makon a matsayin abin da ake kira shirin multimodal, wanda zai iya kunna wasannin bidiyo, hira, rubuta abubuwan da aka tsara, hotuna masu taken, da sarrafa tubalan tara hannun mutum-mutumi. Cibiyar sadarwa ce guda ɗaya wacce zata iya aiki tare da nau'ikan bayanai da yawa don aiwatar da ayyuka iri-iri. 

"Tare da saitin ma'auni guda ɗaya, Gato na iya shiga tattaunawa, hotuna mai taken, tari tare da hannun mutum-mutumi na gaske, ƙwararrun mutane wajen yin wasannin Atari, kewaya cikin yanayin 3D da aka kwaikwaya, bi umarni, da ƙari," marubucin marubuci Scott Reed da abokan aiki a cikin takarda, "A Generalist Agent," wanda aka buga akan sabar preprint Arxiv. 

Demis Hassabis, wanda ya kafa DeepMind, ya yi murna da tawagar, furta a cikin wani tweet, “Babban wakilinmu har yanzu!! Kyakkyawan aiki daga ƙungiyar! " 

Hakanan: Wani sabon gwaji: Shin AI da gaske ya san kuliyoyi ko karnuka - ko wani abu?

Iyakar abin da aka kama shi ne cewa Gato a zahiri bai yi girma sosai akan ayyuka da yawa ba. 

A gefe guda, shirin yana iya yin aiki mafi kyau fiye da shirin koyo na injina wajen sarrafa hannu na Sawyer na mutum-mutumi wanda ke tara tubalan. A gefe guda, yana samar da taken hotuna waɗanda a yawancin lokuta ba su da kyau sosai. Ƙarfinsa a daidaitaccen tattaunawar taɗi tare da mai magana da ɗan adam yana da matsakaicin matsakaici, wani lokacin yana haifar da maganganu masu karo da juna da marasa ma'ana. 

Kuma wasansa na wasannin bidiyo na Atari 2600 ya faɗi ƙasa da na mafi yawan shirye-shiryen ML da aka keɓe don yin gasa a cikin ma'auni. Muhallin Koyon Arcade. 

Me ya sa za ku yi shirin da ke yin wasu abubuwa da kyau da kuma tarin wasu abubuwa marasa kyau? Precedent, da kuma tsammanin, a cewar marubutan. 

Akwai misali don ƙarin nau'ikan shirye-shirye na gabaɗaya sun zama yanayin fasaha a cikin AI, kuma akwai tsammanin cewa ƙara yawan ikon sarrafa kwamfuta zai haifar da gazawa a nan gaba. 

Gabaɗaya na iya ƙoƙarin yin nasara a cikin AI. Kamar yadda marubutan suka lura, suna ambaton masanin AI Richard Sutton, "A tarihi, nau'ikan nau'ikan nau'ikan da suka fi dacewa wajen yin lissafin ƙididdiga suma sun kasance suna ci gaba da ƙetare ƙayyadaddun hanyoyin yanki na musamman daga ƙarshe."

Kamar yadda Sutton ya rubuta a cikin nasa blog post, "Babban darasi da za a iya karantawa daga shekaru 70 na bincike na AI shine cewa hanyoyin gabaɗaya waɗanda ke yin amfani da ƙididdigewa sun kasance mafi inganci, kuma ta wani babban gefe."

Sanya a cikin ƙayyadaddun kasida, Reed da ƙungiyar sun rubuta cewa “mu a nan muna gwada hasashen cewa horar da wakili wanda gabaɗaya yana iya yin ayyuka da yawa yana yiwuwa; kuma za a iya daidaita wannan wakilin gabaɗaya tare da ƴan ƙarin bayanai don yin nasara a mafi girman adadin ayyuka."

Hakanan: LeCun na Meta's AI yana bincika iyakar kuzarin koyo mai zurfi

Samfurin, a cikin wannan yanayin, shine, hakika, gabaɗaya. Sigar Transformer ce, mafi girman nau'in ƙirar tushen kulawa wanda ya zama tushen yawancin shirye-shirye ciki har da GPT-3. Transformer yana ƙirar yuwuwar wani abu da aka ba da abubuwan da ke kewaye da shi kamar kalmomi a cikin jumla. 

A game da Gato, masana kimiyya na DeepMind sun sami damar yin amfani da binciken yuwuwar sharadi iri ɗaya akan nau'ikan bayanai da yawa. 

Kamar yadda Reed da abokan aiki suka bayyana aikin horar da Gato, 

A lokacin horo na Gato, bayanai daga ayyuka daban-daban da hanyoyin ana jera su zuwa jeri-nauyi na alamomi, batches, da sarrafa su ta hanyar hanyar sadarwa ta jijiyoyi masu kama da babban ƙirar harshe. An rufe asarar ta yadda Gato kawai yayi hasashen aiki da hari na rubutu.

Gato, a wasu kalmomi, ba ya bi da alamomi daban-daban ko kalmomi ne a cikin hira ko motsi a cikin motsa jiki. Duk daya ne. 

An binne shi a cikin Reed da hasashe na ƙungiyar, shine cewa ƙara ƙarfin kwamfuta zai yi nasara, a ƙarshe. A halin yanzu, Gato yana iyakance ta lokacin amsawar hannun mutum-mutumi na Sawyer wanda ke yin shingen toshewa. A sigogin cibiyar sadarwa biliyan 1.18, Gato ya fi ƙanƙanta da manyan samfuran AI kamar GPT-3. Yayin da ƙirar ilmantarwa mai zurfi ke girma, yin tunani yana haifar da jinkirin da zai iya gazawa a cikin duniyar da ba ta yanke hukunci ba na ainihin mutum-mutumi. 

Amma, Reed da abokan aiki suna tsammanin za a wuce wannan iyaka yayin da kayan aikin AI ke samun saurin sarrafawa.

"Muna mai da hankali kan horar da mu a wurin aiki na sikelin samfurin wanda ke ba da damar sarrafa na'urori na zamani na ainihi, a halin yanzu a kusa da sigogi na 1.2B a cikin yanayin Gato," sun rubuta. "Yayin da kayan masarufi da kayan gine-ginen ke haɓaka, wannan wurin aiki a zahiri zai haɓaka girman samfurin da zai yuwu, yana tura samfuran gama gari sama da tsarin doka."

Don haka, Gato haƙiƙa abin ƙira ne na yadda ma'aunin ƙididdigewa zai ci gaba da kasancewa babban jigon ci gaban koyan na'ura, ta hanyar samar da samfuran gaba ɗaya girma da girma. Girma ya fi kyau, a wasu kalmomi. 

Kuma marubutan suna da wasu shaidu kan hakan. Gato yana da alama yana samun kyau yayin da yake girma. Suna kwatanta matsakaitan maki a cikin dukkan ayyukan ma'auni na ƙira masu girma uku bisa ga sigogi, miliyan 79, miliyan 364, da babban samfurin, biliyan 1.18. "Za mu iya ganin cewa don daidaitaccen ƙidayar alamar, akwai gagarumin ci gaba tare da ƙara yawan sikelin," marubutan sun rubuta. 

Tambaya mai ban sha'awa a nan gaba ita ce ko shirin da ke da cikakken bayani ya fi hatsari fiye da sauran nau'o'in shirye-shiryen AI. Marubutan suna ciyar da lokaci mai yawa a cikin takarda suna tattauna gaskiyar cewa akwai haɗarin haɗari da ba a fahimta sosai ba tukuna.  

Tunanin shirin da ke gudanar da ayyuka da yawa yana nuna wa ɗan adam nau'in daidaitawar ɗan adam, amma hakan na iya zama mummunar fahimta. "Alal misali, yanayin jiki zai iya haifar da masu amfani anthropomorphizing wakili, haifar da rashin amincewa ga yanayin tsarin da ba daidai ba, ko kuma yin amfani da shi ta hanyar miyagun 'yan wasan kwaikwayo," Reed da tawagar sun rubuta. 

"Bugu da ƙari, yayin da canja wurin ilimin yanki sau da yawa manufa ce a cikin binciken ML, yana iya haifar da sakamakon da ba a zata ba idan aka canza wasu halaye (misali faɗar wasan arcade) zuwa mahallin da ba daidai ba."

Don haka, sun rubuta, "Tsarin da'a da aminci na canja wurin ilimi na iya buƙatar ingantaccen bincike yayin da tsarin tsarin gaba ɗaya ya ci gaba."

(A matsayin bayanin kula mai ban sha'awa, takardar Gato ta yi amfani da wani tsari don kwatanta hadarin da tsohon mai bincike na Google AI Margaret Michell da abokan aiki suka tsara, wanda ake kira Model Cards. Katin Model yana ba da taƙaitaccen taƙaitaccen abin da shirin AI yake, abin da yake aikatawa, da abin da yake aikatawa. A bara, Michell ta rubuta cewa an tilasta mata ficewa daga Google saboda goyon bayan tsohon abokin aikinta, Timnit Gebru, wanda damuwarsa game da AI ya ci karo da jagorancin Google na AI.)

Gato ko kaɗan ba shi da wata maɓalli a cikin halinsa na gaba ɗaya. Yana daga cikin faɗuwar yanayin gabaɗaya, da manyan samfuran da ke amfani da bokitin dawakai. Duniya ta sami ɗanɗano na farko na karkatar da Google a wannan lokacin bazarar da ta gabata, tare da Google's “Perceiver” neuron network wanda ya haɗa ayyukan Transformer na rubutu tare da hotuna, sauti, da daidaitawar sararin samaniya na LiDAR.

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

Daga cikin takwarorinta akwai PaLM, Samfurin Harshen Hanyoyi, Masana kimiyyar Google ne suka gabatar da wannan shekarar, samfurin siga na biliyan 540 wanda ke yin amfani da sabuwar fasaha don daidaita dubban kwakwalwan kwamfuta, aka sani da Pathways, kuma an ƙirƙira shi a Google. Cibiyar sadarwa ta jijiyar da Meta ta fitar a watan Janairu, wanda ake kira "data2vec," tana amfani da Transformers don bayanan hoto, sautin murya na magana, da wakilcin harshen rubutu duk a daya. 

Abin da ke sabo game da Gato, da alama, shine niyyar ɗaukar AI da aka yi amfani da ita don ayyukan da ba na robotic ba da tura shi cikin duniyar robotics.

Masu kirkiro Gato, suna lura da nasarorin Hanyoyi, da sauran hanyoyin gabaɗaya, suna ganin babban nasara a cikin AI wanda zai iya aiki a cikin ainihin duniya, tare da kowane nau'in ayyuka. 

"Ya kamata aikin nan gaba ya yi la'akari da yadda za a haɗa waɗannan damar rubutun zuwa cikakken wakili na gaba ɗaya wanda kuma zai iya aiki a ainihin lokacin a cikin ainihin duniya, a cikin yanayi daban-daban da siffofi." 

Kuna iya, don haka, la'akari da Gato a matsayin muhimmin mataki akan hanyar magance matsalar AI mafi wahala, robotics.