1-Bit da Ternary BitNet Model
BitNet shine layin bincike na Microsoft yana nuna cewa ana iya horar da manyan nau'ikan harshe tare da iyakance ma'auni zuwa kawai 1 bit, ko ƙima uku a cikin harka ta ternary.
Dubawa
This slashes memory and energy use dramatically while keeping surprisingly strong accuracy.
Zurfafa nutsewa
Samfuran al'ada suna adana kowane nauyi azaman lamba 16-bit. BitNet ya maye gurbin waɗannan tare da matsananci ƙananan wakilci. Bambancin BitNet b1.58 mai tasiri yana amfani da ma'auni na uku, kowanne an iyakance shi zuwa -1, 0, ko +1, wanda ke aiki zuwa kusan 1.58 rago na bayanai akan kowane nauyi (log tushe 2 na 3). Muhimmin ra'ayi shine cewa ƙirar an horar da ita daga karce tare da waɗannan ƙuntatawa, ba a ƙididdige su daga baya ba, don haka ya koyi zama mai ƙarfi zuwa ƙayyadaddun daidaito. Saboda ma'auni ne kawai -1, 0, ko +1, tsada mai yawa a cikin matrix math yana rushewa zuwa ƙari da raguwa. Sakamakon yana da nisa ƙananan bandwidth na ƙwaƙwalwar ajiya, amfani da makamashi, da latency, tare da ƙimar 0 kuma yana ba da damar ɓarna, duk yayin da ya dace da ingantattun samfura a girman kwatankwacin kan ma'auni da yawa.
Fahimtar Fasaha
BitNet yana amfani da Layer BitLinear na al'ada wanda ke ƙididdige ma'auni zuwa na uku da kunnawa zuwa ƙananan daidaito yayin wucewar gaba, yayin da yake kiyaye kwafin 'inuwa' mafi girma na ma'aunin nauyi don sabuntawar gradient ta hanyar ƙididdigewa kai tsaye. Saboda kowane nauyi shine -1, 0, ko +1, samfuran ɗigo waɗanda ke mamaye lissafin taswira sun zama ƙari da ragi maimakon ninka-maki-ruwa, wanda shine ke buɗe ƙarfin kuzari da ribar sauri akan kayan aikin da suka dace.
Dabarun Tasiri
Kudin da kasafin kuɗi
Hukunce-hukuncen gine-gine suna haifar da aiki da tsadar aiki na shekaru.
Shawarwari masu haske
Ilimin fasaha yana taimaka wa ƙungiyoyi su zaɓi tari mai kyau, ba kawai sabon abu ba.
Kula da inganci
Zaɓuɓɓukan injiniya mafi kyau suna rage abin dogaro a cikin samarwa.
Makomar 1-Bit da Tsarin BitNet na Ternary
BitNet yana nuni zuwa gaba inda samfura masu ƙarfi ke gudana akan wayoyi, kwamfyutocin kwamfyutoci, da na'urorin gefen ba tare da GPUs na datacenter ba. Babban ƙwanƙwasa kayan aiki ne: kwakwalwan kwamfuta na yau an gina su ne don lissafi mai iyo, don haka ƙwararrun masu haɓakawa waɗanda aka inganta don ƙarin ayyuka na ternary-kawai na iya haɓaka fa'idodin. Yi tsammanin ƙarin ƙirar gine-ginen 1-bit na ƙasa, samfuran tsarin BitNet mafi girma, da haɗin kai cikin mataimakan na'urori inda rayuwar baturi da keɓantawa, mai yuwuwar sake fasalin tattalin arziƙin AI.
Aiwatar da Gaskiyar Duniya
Microsoft's BitNet b1.58 2B4T yana gudana yadda ya kamata akan CPU, yana ba da damar tantance LLM ba tare da keɓaɓɓen GPU ba.
Mataimakan kan na'ura waɗanda suka dace da ƙirar ƙira cikin ƙayyadaddun ƙwaƙwalwar ajiyar waya godiya ga ma'aunin ~1.58-bit.
Rage ƙimar ƙima da ƙimar carbon don sabis na API mai girma ta hanyar maye gurbin abubuwan da ke iyo ninkawa tare da ƙari.
Aiwatar da Edge (IoT, kayan aikin da aka haɗa) inda ma'aunin nauyi ya sa fahimtar harshen gida ya yiwu a cikin matsananciyar kasafin kuɗi.
Hatsari & Tsare-tsare
Haɓaka ma'auni ɗaya na iya ɓoye manyan raunin tsarin.
Sau da yawa ana raina kayan more rayuwa da kuma kuɗin kulawa.
Tsaro da gibin lura na iya girma yayin da tsarin ke ƙara haɓaka.
Taswirar Hanya
Ƙayyade latency, inganci, da maƙasudin farashi kafin aiwatarwa.
Alamar ma'auni a ƙarƙashin ainihin kaya da yanayin bayanai.
Kula da kayan aiki don kurakurai, ɗigo, da tasirin mai amfani.
Shirya bijirowa da hanyoyin mayar da martani kafin sikeli.
Ci gaba da Bincike
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the 1-Bit and Ternary BitNet Models quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Jagora na gaba
Abubuwan Samfuran Harshen Watermarking
Tambayoyin da ake yawan yi
What is 1-Bit and Ternary BitNet Models?
BitNet shine layin bincike na Microsoft yana nuna cewa ana iya horar da manyan nau'ikan harshe tare da iyakance ma'auni zuwa kawai 1 bit, ko ƙima uku a cikin harka ta ternary. Wannan yana rage ƙwaƙwalwar ajiya da amfani da kuzari sosai yayin kiyaye daidaito mai ƙarfi mai ban mamaki.
Me yasa BitNet b1.58 aka kwatanta da amfani da kusan 1.58 rago a kowace nauyi?
Ma'aunin nauyi na ƙasa yana ɗaukar ɗaya daga cikin ƙima guda uku, kuma wakiltar jihohi uku yana buƙatar tushen log 2 na 3, kusan 1.58 rago.
Me yasa ayyukan matrix na BitNet ya fi arha fiye da daidaitattun samfura?
Tare da iyakance ma'auni zuwa -1, 0, da +1, ninkawa ta nauyi yana ragewa zuwa ƙarawa, ragi, ko watsi da ƙima, guje wa ƙima mai tsadar gaske.
Yayin horo, ta yaya BitNet ke sabunta ma'auni duk da matakin ƙididdigewa wanda ba shi da bambanci?
BitNet yana kiyaye kwafin ma'aunin nauyi mafi girma kuma yana amfani da madaidaicin-ta hanyar ƙididdigewa don wuce gradients ta hanyar ƙididdigewa, yana ba da damar yaɗa baya na yau da kullun.
Wane ƙarin fa'ida ke ba da izinin ƙimar 0 (ternary, ba tsantsar binary) ba?
Jihar 0 tana ba da damar kashe wasu hanyoyin sadarwa yadda ya kamata, suna gabatar da ɓacin rai waɗanda za a iya amfani da su don ƙarin inganci.
Menene babban ƙalubalen kayan masarufi don fahimtar cikakken nasarar BitNet?
An ƙera na'urori na yau da kullun a kusa da ninka-matsakaici, don haka keɓaɓɓen kayan aiki don ayyukan tushen ƙari na uku ana buƙatar don buɗe cikakkiyar fa'idodin saurin BitNet da kuzari.