Slav Petrov
e-@ Google DeepMind
Vice President, Research at Google DeepMind; co-leads Post-Training for Gemini
EJamani Themes
USlav Petrov ngumcwaningi waseBulgaria kunye noSekela Mongameli wezePhando kwiGoogle DeepMind, apho akhokela khona ngokubambisana uQeqesho-emva (Post-Training) kwaGemini, umzamo kaGoogle wemodeli yolwimi emikhulu. Unayo iPhD kwiYunivesithi yaseCalifornia, eBerkeley, kwaye ufundise iStatistical Natural Language Processing kwiYunivesithi yaseNew York.
Uphando lwakhe lugxile ekucutshungulweni kolwimi lwendalo nakwiimodeli zolwimi ezinkulu, kwaye umsebenzi wakhe uye wanceda ekunikezeni iGoogle iindlela zokusebenza ngolwimi kwiimveliso ezifana neSearch, Assistant, Translate, neCloud. Ufumene iiMbasa ezininzi zeBest Paper kwiinkomfa eziphambili ze-NLP (ACL 2011, NAACL 2012, ACL 2016), kwakunye neMbasa yeTest-of-Time yeminyaka eli-10 kwi-ACL 2023, kwaye wanikwa iJohn Atanasoff Award ngo-2014 nguMongameli waseBulgaria.
Imvelaphi Yezifundo Nezobuchwephesha
Uhambo lwezifundo lukaPetrov luthatha amazwe amathathu kwaye lubonisa ubume bamazwe ngamazwe bomsebenzi wakhe. Waqhuba iDiplom kwiComputer Science neBusiness Administration eFreie Universität Berlin ngaphambi kokuba aqhubeke nezifundo zobugqirha kwiComputer Science kwiYunivesithi yaseCalifornia, Berkeley. Kamva wafunda eDuke University njengenxalenye yendlela yakhe yangaphambili yezifundo. Ukusukela ngo-2010 ukuya ku-2016, wasebenza njengeAdjunct Professor kwiNew York University, apho wafundisa iStatistical Natural Language Processing — indima eyayiqhutywa ngaxeshanye nomsebenzi wakhe weshishini eGoogle, apho waqala khona ngoAgasti 2009.
Indima eGoogle DeepMind
EGoogle DeepMind, uPetrov unesikhundla seSekela Mongameli weZifundo (Vice President of Research) kwaye uyahola ngokudibeneyo iPost-Training yeGemini, imodeli enkulu yolwimi ephambili yenkampani. IPost-training yinqanaba elibalulekileyo ekuphuhlisweni kweemodeli zolwimi ezinkulu, elibandakanya iindlela ezilolonga indlela imodeli eziphatha ngayo, amandla okulandela imiyalelo, kunye nokuhambelana (alignment) emva kokugqitywa kokuqala koqeqesho (pretraining). Iqela lakhe lisasazwe emhlabeni jikelele, nto leyo ebonisa ubungakanani nobunzima bomsebenzi obandakanyekayo. Ngaphambi kokusekwa kweGoogle DeepMind, uPetrov wayehola iinzame zophando zeNLP ezazingena ngokuthe ngqo kwiimveliso zeGoogle ezifana neSearch, Ads, Translate, Assistant, neChrome.
Ingxelo yakhe yokushicilela ihlanganisa ngaphezulu kweminyaka elishumi kwaye igubungela uluhlu olubanzi lwezihloko ngaphakathi kwe-natural language processing, kubandakanya ukuguqulelwa komatshini, i-parsing, kunye ne-attribution ekwenzeni ulwimi ngokwendalo. Ungumxhasi kumaphepha aphawulekayo ahambelana neenzame zeGoogle zemodeli enkulu, kuquka umsebenzi kwiPaLM, PaLM 2, Gemini, Gemini 1.5, neGemma 2, kunye nedathasethi ye-benchmark yeNatural Questions — isixhobo esisetyenziswa kakhulu kuphando lokuphendula imibuzo.
Igalelo loPhando nokuQatshelwa
Imveliso yophando lukaPetrov ibandakanya igalelo kwi-syntactic parsing, ukuguqulelwa komatshini, kunye nokuphuhliswa kweendlela zokuvavanya (evaluation frameworks) ekwenzeni ulwimi. Enye indawo ephawulekayo yomsebenzi wakhe wamva nje ibandakanya ukulinganisa i-attribution kwiimodeli zokwenziwa kolwimi — ukujonga ukuba imveliso yemodeli inokulandwa iye kwimithombo enokuthenjwa, inkxalabo eye yanda ukubaluleka kwayo ecaleni kokusetyenziswa kweenkqubo ezinkulu zokwenziwa kolwimi. Lo msebenzi, opapashwe kwiComputational Linguistics ngo-2023, ubonisa umdla obanzi ekwenzeni imveliso yeelwimi zeemodeli ithembeke ngakumbi kwaye icaciseke.
Ngaphandle kwamabhaso akhe ephepha kwi-ACL ne-NAACL, uPetrov wafumana i-10-year Test-of-Time Award kwi-ACL 2023, ekhomba uphando olwalusenziwe kwiminyaka elishumi ngaphambili olungazange luphelelwe lixesha kwaye lwabonakala luhlala lunempembelelo kulo mmandla. Iindawo zakhe zophando, njengoko zidweliswe kwindawo yophando kaGoogle, zigubungela iNatural Language Processing, Machine Translation, Machine Intelligence, neGeneral Science.
Imvelaphi nezinye iimpumelelo
U-Petrov uhlala eBulgaria, eJamani, nase-United States ngokulinganayo okukhulu, kwaye ngoku useBerlin. Iimpumelelo zakhe zokuqala zokhuphiswano ziquka iWorld Championship eRoboCup 2004 — ukhuphiswano lwamanye amazwe lweerobhothi — olwandulela umsebenzi wakhe wophando lwe-NLP. Waqhuba iDiplom yakhe eFreie Universität Berlin, kwaye ixesha lakhe eJamani libonakala limi njengencwadi eqingqiweyo yomsebenzi wakhe, kunye neBerlin ibe sisiseko sakhe sangoku nangona ehola uphando kwenye yezona mibutho ye-AI idume kakhulu emhlabeni. Uye wavakalisa ukuthembeka kwakhe komntu kwiqela lebhola ekhatywayo lesizwe laseBulgaria, ngakumbi kwimidlalo ngokuchasene neJamani.