da Masu binciken Google sun saki labarai game da ci gaban sabon tsarin wanda ya faɗaɗa horar da ƙirar ilimin kere kere zuwa dubunnan inji. Ana kiran sakamakon GABA RL (ƙwarewar ingantaccen ilmantarwa mai ƙarfi).
Wannan shi ne alamar cigaba saboda yakamata ba da damar koyar da ilimin kere-kere na kere-kere zuwa miliyoyin hotuna a dakika daya kuma rage farashin wannan horon da kashi 80%, in ji Google a cikin takardar bincike.
Irin wannan rage girman zai iya taimakawa daidaita filin wasan don farawa. cewa har zuwa yanzu ba su iya yin gasa tare da manyan kamar Google a fagen AI ba. Kudin horar da samfuran ilmantarwa na zamani a girgije abun mamaki ne. Google ya ƙaddamar da buɗe lambar SEED RL, aikin da aka tsara don haɓaka ƙimar farashi / aikin haɓaka ƙarfin ilmantarwa.
Ilimin karfafa gwiwa hanya ce takamaimai wacce ake amfani da ita - inda wakilai ke koyo game da muhallin su ta hanyar bincike da kuma inganta ayyukansu don mafi yawan lada.
A cikin »SEED RL: Mai iya aunawa da Ingantaccen Deep-RL tare da Ingantaccen Centraladdamarwa ta Tsakiya," mun gabatar da wakili na RL wanda ke auna wa dubunnan injuna, wanda ke ba da horo a miliyoyin sigogi a kowane dakika kuma yana inganta ingantaccen lissafi. Ana samun wannan tare da tsarin gine-ginen labari wanda ke amfani da hanzari (GPU ko TPU) a sikeli ta hanyar sanya ƙirar samfuri da gabatar da hanyar sadarwa mai sauri.
Muna nuna aikin SEED RL akan sanannun alamun RL kamar Google Research Football, Arcade Learning Environment, da DeepMind Lab, kuma muna nuna cewa ta amfani da manyan siloli, ƙwarewar bayanai na iya ƙaruwa. An buɗe lambar a kan Github tare da misalai don gudana akan Google Cloud tare da GPU.
SEED RL ya dogara ne akan tsarin TensorFlow 2.0 y yana aiki ta amfani da haɗin ɗakunan sarrafa hoto da sassan sarrafa tensor don karkatar da tsarin samfura. Ana yin amfani da hankali a tsakiya ta amfani da ɓangaren ilmantarwa wanda ke horar da ƙirar.
Ana adana masu canji da bayanan jihar na ƙirar ƙirar gida kuma ana aika musu da lura akan su a kowane mataki na aikin. SEED RL kuma yana amfani da laburaren cibiyar sadarwa bisa tushen tsarin buɗe RPC na duniya don rage latenci.
da Masu binciken Google sun ce bangaren ilmantarwa by Tsakar Gida za a iya fadada zuwa dubunnan maɗaukaki, yayin da yawan 'yan wasan da za a maimaita su tsakanin ɗaukar ma'auni a cikin muhalli da aiwatar da tunani a kan samfurin don hango abin da ke gaba, ana iya fadada har zuwa dubban injuna.
Google ya kimanta tasirin SEED RL ta hanyar kwatanta shi da sanannen yanayin ilmantarwa na Arcade, yanayin binciken ƙwallon ƙafa na Google, da mahalli daban-daban na DeepMind Lab. ta amfani da kwakwalwan komputa guda 2,4 na na'urar sarrafa firikwensin girgije.
Google ya ce ya yi saurin kusan sau 80 fiye da na baya.
"Wannan yana fassara cikin mahimmancin hanzarin lokaci, kamar yadda masu hanzari ke da rahusa ta kowane aiki fiye da CPUs, farashin abubuwan gwaje-gwaje ya ragu sosai." Mun yi imanin SEED RL kuma sakamakon da aka gabatar ya nuna cewa karantar da ilmantarwa ya sake riskar sauran ilmin mai zurfin dangane da amfani da hanzari, "in ji Lasse Espeholt, injiniyan bincike a Google Research.
Tare da gine-ginen da aka inganta don amfani dashi cikin hanzari na zamani, abu ne na al'ada don ƙara girman ƙirar a yunƙurin haɓaka ƙimar bayanai.
Google ya ce lambar SEED RL ta kasance budaddiyar hanya ce kuma ana samun ta a Github, kazalika da misalai da ke nuna yadda ake samun sa ta yin aiki a kan Google Cloud tare da sassan sarrafa hotuna.
Aƙarshe, ga waɗanda suke da sha'awar wannan sabon tsarin, zasu iya zuwa mahaɗin mai zuwa inda zasu sami ƙarin bayani game da shi. Haɗin haɗin shine wannan.
Source: https://ai.googleblog.com/