Google DeepMind Princeton

Dikgokagano tsa Patlisiso ya Ithutopalo ya Thuto ya Mechini

Google DeepMind Princeton e dira jaaka lefelo la patlisiso le le kgethegileng le le lebisitsweng mo motheong wa dipalo wa thuto ya mechini. Sehlopha se tlhabolola ditsamaiso tse di leka-lekanang tshebetso e e dirang sentle le dipoelo tsa mo tlhaloganyong ya dipalo, go netefatsa go tiya mo tirisong ya mmatota.

Dikarolo tsa botlhokwa tsa patlisiso di akaretsa spectral transformers, adaptive gradient methods, le non-stochastic control theory. Ditsela tseno di ikaeletse go tokafatsa bokgoni mo go ruteng diswantšho (models) fa di ntse di boloka go tlhaloganyega le go ikanyega.

Mo tirong ya lefelo, go tsweledisiwa gape mo dikagong tsa kalafi, jaaka go tokafatsa ditsamaiso tsa taolo ya phepelo ya moya (ventilation). Se se bontsha gore diphetogo tsa mo tlhaloganyong ya dipalo di ka fetoga go nna ditokafatso tse di bonalang mo thekenoloji ya tlhokomelo ya botsogo.

Dikgokagano tsa Tšhupano le tsa Thuto

Lefelo la Princeton le na le kamano e e gaufi le Yunibesiti ya Princeton, le tlhama tikologo e e kopanyang bokgoni jwa thuto le boitlhamelo jwa intaseteri. Kwa e leng teng gaufi le khampase go dira gore go nne le phapanyetsano e e sa kgaotseng le barutabana le baithuti, go humisa patlisiso le thuto.

Dikgogakano di ralala ditlhopha di le mmalwa tsa patlisiso tsa Google mo U.S., India, Israel, le Yuropa. Netweke eno ya lefatshe lotlhe e letlelela lefelo go kopanya maikutlo a a farologaneng le go potlakisa tswelelopele mo patlisisong ya motheo ya AI.

Bokgolo jwa sehlopha bo tswelela bo le kwa tlase ka boomo, bo etelela pele botebo go na le bogolo. Sebopego seno se tshegetsa go batlisisa ka kelotlhoko mathata a a raraaneng fa se ntse se boloka bokgoni jwa go fetola tsela ya patlisiso.

Diprojeke tsa Bohlokwa le Diphetogo

Diprojeke tsa lefelo di bontsha go kopanngwa ga patlisiso ya dipalo le e e dirisiwang. Spectral transformers, ka sekai, di batlisisa mekgwa e mešwa ya go aga dikarolo tsa go rulaganya tatelano (sequence modeling), fa adaptive gradient methods tsona di tokafatsa ditsela tsa go tlhabolola (optimization) mo go ruteng ka bogolo jo bo kwa godimo mo thutong ya mechini.

Mo go ithuteng ka go itshwara (reinforcement learning), sehlopha se tlhophile ditsela tse di ka tlhaloganyegang (differentiable) le non-stochastic control theory. Diphetogo tseno di rarabolola mathata mo tikologong e e fetogang, kwa ditsela tsa setso di sa lekaneng.

Dikopo tsa kalafi di tswelela e le kgatiso e kgolo. Morero wa AI for Better Medical Ventilators o dirisa thuto ya mechini go tokafatsa dipoelo tsa bakuli mo maemong a tlhokomelo e e botlhokwa. Ditiro tse dingwe, jaaka SAMUEL le GGT, di lebile mo go tlhaboleng ga adaptive regularization ka tsela e e bolokang kgopolo (memory-efficient) go tshegetsa go ruta ka bogolo jo bo kgonagalang.

Sehlopha le Tšusumetso ya Patlisiso

Lefelo le etelelwa pele ke baitsesa ba ba nang le bokgale mo saense ya khomputara, optimization, le control theory. Ba le teng jaanong ba akaretsa Elad Hazan, Wenhan Xia, le Ani Majumdar, mo gare ga ba bangwe, ka bokgoni jo bo ralala go rulaganya ditsamaiso, reinforcement learning, le AI ya kalafi.

Ba pele ba tlogetse ba fetogile go nna ditiro mo thutong le mo intasetering, ba tsenya letsogo mo tswelepeleng e e pharaletseng mo thutong ya mechini. Ditiro tsa bone di gantsi di bonala mo dikopanong tse di kwa godimo, jaaka NeurIPS le ICML, go bontsha nonofo ya lefelo mo lebaleng.

Ditsamaiso (publications) tse di tswang mo lefelong di akaretsa mefuta ya ditlhogo, go tswa mo go ithuteng go sa lekanyeng ga nonlinear dynamics go ya mo go lekanyeng (benchmarking) maano a roboto. Ditiro tsa bosheng di akaretsa go lekola dikai tse dikgolo tsa puo (large language models) mo tikologong e e etsisitsweng, le go tlhabolola mekgwa ya polokesego (safety frameworks) ya ditsamaiso tsa roboto.

Patlisiso ya lefelo e tswelela gape mo go ruteng ditsamaiso, ka diprojeke tse jaaka VeLO tse di godisang (scaling) optimizers tse di ithutilweng go di dirisa mo go pharaletseng. Diteko tseno di gatelela boitlamo jwa sehlopha jwa go kopanya tlhaloganyo ya dipalo le tiriso mo tlhabololong ya AI.