Keep Calm and Stay Smart
04:29
Meta's AI guru LeCun: Feem ntau niaj hnub no txoj hauv kev AI yuav tsis ua rau muaj kev txawj ntse tseeb
Yann LeCun, Meta tus kws tshawb fawb AI hais tias "Kuv xav tias AI systems yuav tsum muaj peev xwm ua kom muaj laj thawj," Niaj hnub no cov AI nrov tshaj plaws xws li Transformers, ntau yam uas tsim los ntawm nws tus kheej pioneering ua hauj lwm nyob rau hauv lub tshav pob, yuav tsis txaus. LeCun hais tias, "Koj yuav tsum tau nqis ib kauj ruam rov qab thiab hais tias, Okay, peb ua tus ntaiv no, tab sis peb xav mus rau lub hli, thiab tsis muaj ib txoj hauv kev uas tus ntaiv no yuav mus rau peb tau," LeCun hais.
Yann LeCun, tus thawj kws tshawb fawb AI ntawm Meta Properties, tus tswv ntawm Facebook, Instagram, thiab WhatsApp, zoo li yuav zuam tawm ntau tus neeg hauv nws daim teb.
Nrog kev tshaj tawm hauv Lub Rau Hli ntawm ib qho kev xav ntawm Open Review server, LeCun muab cov ntsiab lus dav dav ntawm txoj hauv kev uas nws xav tias tuav cov lus cog tseg rau kev ua tiav tib neeg qib kev txawj ntse hauv tshuab.
Nrov tam sim no
Qhov cuam tshuam yog tias tsis hais hauv daim ntawv yog qhov kev sib cav uas feem ntau ntawm cov phiaj xwm loj niaj hnub no hauv AI yuav tsis muaj peev xwm ncav cuag tib neeg lub hom phiaj.
Hauv kev sib tham lub hli no nrog ZDNet ntawm Zoom, LeCun tau qhia meej tias nws pom nrog kev tsis ntseeg siab ntau ntawm txoj hauv kev zoo tshaj plaws ntawm kev tshawb fawb hauv kev kawm tob hauv lub sijhawm tam sim no.
"Kuv xav tias lawv tsim nyog tab sis tsis txaus," Turing Award tus yeej tau hais ZDNet ntawm nws cov phooj ywg ' pursuits.
Cov no suav nrog cov qauv lus loj xws li Transformer-based GPT-3 thiab lawv cov ilk. Raws li LeCun tus cwj pwm nws, Transformer devotées ntseeg, "Peb tokenize txhua yam, thiab cob qhia loj heev.cov qauv los ua qhov kev kwv yees tsis sib xws, thiab qee qhov AI yuav tawm ntawm qhov no. "
Nws hais tias, "Lawv tsis yog qhov yuam kev," nws hais tias, "nyob rau hauv qhov kev nkag siab tias tej zaum yuav yog ib feem ntawm lub neej yav tom ntej ntse, tab sis kuv xav tias nws yog ib qho tseem ceeb uas ploj lawm."
Tsis tas li ntawd: Meta's AI luminary LeCun tshawb nrhiav kev kawm tob txog lub zog pem hauv ntej
Nws yog qhov tsis txaus ntseeg ntawm qhov tshwm sim ua haujlwm los ntawm cov kws tshawb fawb uas ua tiav kev siv cov kev sib txuas lus neural, cov txheej txheem siv tau zoo kawg nkaus hauv cov kev kawm tob.
LeCun pom qhov tsis zoo thiab kev txwv hauv ntau qhov chaw ua tau zoo ntawm kev qhuab qhia.
Kev txhawb zog kev kawm kuj yuav tsis txaus, nws tuav. Cov kws tshawb fawb xws li David Silver ntawm DeepMind, uas tau tsim lub AlphaZero program uas tau paub Chess, Shogi thiab Go, tab tom tsom mus rau cov kev pab cuam uas "ua haujlwm zoo heev," saib LeCun, tab sis "feem ntau ntawm kev kawm peb ua, peb tsis ua. ua nws los ntawm kev ua tiag tiag, peb ua nws los ntawm kev soj ntsuam. "
Lecun, 62, los ntawm qhov kev xav ntawm ntau xyoo ntawm kev ua tiav, txawm li cas los xij qhia qhov kev kub ntxhov los tawm tsam qhov nws xav tias yog cov qhov muag tsis pom kev mus rau qhov uas ntau tus yuav maj nrawm, thiab sim ua kom nws lub teb nyob rau hauv kev coj nws xav tias txhua yam yuav tsum mus.
Nws tau hais tias "Peb pom ntau qhov kev thov hais tias peb yuav tsum ua dab tsi los thawb mus rau tib neeg qib AI," nws hais. "Thiab muaj cov tswv yim uas kuv xav tias yog yuam kev."
Lecun hais tias "Peb tsis mus txog qhov twg peb lub tshuab ntse muaj kev nkag siab zoo ib yam li miv," Lecun hais. "Yog li, vim li cas peb tsis pib muaj?"
Nws tau tso tseg nws txoj kev ntseeg ua ntej hauv kev siv cov khoom siv sib txuas hauv cov khoom xws li kwv yees tus ncej tom ntej hauv video. "Nws tau ua tsis tiav," nws hais.
LeCun txiav txim siab rau cov uas nws hu ua "kev ntseeg txog kev ntseeg," uas "xav tias qhov kev xav txog qhov yuav tshwm sim yog tib lub ntsiab lus uas koj tuaj yeem siv los piav qhia txog kev kawm tshuab."
Nws tau hais tias txoj hauv kev tsuas yog kev txheeb cais yog qhov tsis txaus ntseeg. “Nws yog qhov ntau dhau los thov kom tus qauv hauv ntiaj teb yuav tsum tau ua tiav; peb tsis paub yuav ua li cas. "
Tsis yog cov kev kawm nkaus xwb, tab sis kev lag luam AI xav tau kev rov xav dua, sib cav LeCun. Cov neeg tsav tsheb tus kheej, kev pib ua lag luam xws li Wayve, tau "zoo siab me ntsis," nws hais tias, los ntawm kev xav tias lawv tuaj yeem "tshem cov ntaub ntawv ntawm" cov network neural loj "thiab koj tuaj yeem kawm tau ntau yam."
"Koj paub, kuv xav tias nws muaj peev xwm tag nrho uas peb yuav muaj qib tsib lub tsheb tsis muaj kev paub zoo," nws hais, hais txog "ADAS," advanced driver help system cov ntsiab lus rau kev tsav tsheb tus kheej, "tab sis koj yuav tsum tau tsim lub ntuj raug txim tawm ntawm nws."
Xws li kev tsav tsheb tus kheej dhau los yuav yog ib yam uas creaky thiab nkig raws li tag nrho cov khoos phis tawj pom kev uas tau tsim los ntawm kev kawm tob, nws ntseeg.
"Thaum kawg, yuav muaj kev txaus siab dua thiab muaj peev xwm daws tau zoo dua uas suav nrog cov tshuab ua haujlwm zoo dua los nkag siab txog txoj hauv kev ua haujlwm hauv ntiaj teb."
Nrov tam sim no
Raws li txoj kev, LeCun muaj qee qhov kev xav tsis txaus ntseeg ntawm nws cov neeg thuam loj tshaj plaws, xws li NYU tus xibfwb Gary Marcus - "nws tsis tau pab dab tsi rau AI" - thiab Jürgen Schmidhuber, tus thawj coj ntawm Dalle Molle Institute for Artificial Intelligence Research - "nws yog. yooj yim heev los ua tus chij-cog. "
Tshaj li qhov kev thuam, qhov tseem ceeb tshaj plaws los ntawm LeCun yog tias qee qhov teeb meem tseem ceeb tawm tsam txhua yam ntawm AI, tshwj xeeb, yuav ntsuas cov ntaub ntawv li cas.
"Koj yuav tsum tau nqis ib kauj ruam rov qab thiab hais tias, Okay, peb ua tus ntaiv no, tab sis peb xav mus rau lub hli, thiab tsis muaj txoj hauv kev uas tus ntaiv no mus rau peb tau," LeCun hais ntawm nws lub siab xav kom rov xav txog. ntawm cov tswv yim yooj yim. "Qhov tseem ceeb, qhov kuv sau ntawm no yog, peb yuav tsum tsim cov foob pob hluav taws, Kuv tsis tuaj yeem muab cov ntsiab lus ntawm peb tsim foob pob hluav taws li cas, tab sis ntawm no yog cov hauv paus ntsiab lus."
Daim ntawv, thiab LeCun cov kev xav hauv kev xam phaj, tuaj yeem nkag siab zoo dua los ntawm kev nyeem LeCun qhov kev xam phaj ua ntej xyoo no nrog ZDNet nyob rau hauv uas nws sib cav rau lub zog-raws li tus kheej saib xyuas kev kawm raws li txoj hauv kev rau kev kawm tob. Cov kev xav no ua rau muaj kev nkag siab ntawm qhov tseem ceeb ntawm txoj hauv kev rau qhov nws vam tias yuav tsim los ua lwm txoj hauv kev uas nws hais tias yuav tsis ua rau nws mus txog qhov kawg.
Dab tsi hauv qab no yog ib daim ntawv hloov kho me ntsis ntawm kev xam phaj.
ZDNet: Lub ntsiab lus ntawm peb cov kev sib tham yog daim ntawv no, "Ib txoj hauv kev mus rau kev txawj ntse tshuab kev txawj ntse," ntawm qhov twg version 0.9.2 yog qhov txuas ntxiv, yog?
Yann LeCun: Yog, kuv xav txog qhov no, hom, ib daim ntawv ua haujlwm. Yog li, kuv muab nws tso rau ntawm Qhib Kev Ntsuam Xyuas, tos cov neeg los tawm tswv yim thiab cov lus qhia, tej zaum ntxiv cov ntaub ntawv siv, thiab tom qab ntawd kuv mam li tsim ib qho kev hloov kho.
Nrov tam sim no
ZDNet: Kuv pom tias Juergen Schmidhuber twb tau ntxiv qee cov lus rau Qhib Rov Xyuas.
YL: Zoo, yog, nws ib txwm ua. Kuv sau ib qho ntawm nws cov ntawv nyob rau hauv kuv daim ntawv. Kuv xav tias cov lus sib cav uas nws tau ua hauv kev sib tham uas nws tau tsim txhua yam ntawm no hauv 1991, raws li nws tau ua nyob rau hauv lwm qhov xwm txheej, tsis yog li ntawd. Kuv txhais tau tias, nws yooj yim heev uachij-cog, thiab rau, zoo-ntawm, sau ib lub tswv yim yam tsis muaj kev sim, tsis muaj kev xav, tsuas yog qhia tias koj tuaj yeem ua li no. Tab sis, koj paub, muaj qhov sib txawv loj ntawm tsuas yog muaj lub tswv yim, thiab tom qab ntawd tau txais nws los ua haujlwm ntawm qhov teeb meem khoom ua si, thiab tom qab ntawd tau txais nws los ua haujlwm ntawm qhov teeb meem tiag tiag, thiab tom qab ntawd ua qhov kev xav uas qhia tias vim li cas nws thiaj ua haujlwm, thiab tom qab ntawd. siv nws. Muaj tag nrho cov saw, thiab nws lub tswv yim ntawm scientific credit yog tias nws yog thawj tus neeg uas tsuas yog, xaiv-ntawm, koj paub, muaj lub tswv yim ntawm qhov ntawd, uas yuav tsum tau txais tag nrho cov credit. Thiab qhov ntawd yog qhov tsis txaus ntseeg.
ZDNet: Tsis txhob ntseeg txhua yam koj hnov hauv social media.
YL: Kuv txhais tau tias, daim ntawv tseem ceeb uas nws hais tias kuv yuav tsum hais tsis muaj ib lub tswv yim tseem ceeb uas kuv tham hauv daim ntawv. Nws tau ua qhov no nrog rau GANs thiab lwm yam, uas tsis yog qhov tseeb. Nws yog ib qho yooj yim los ua tus chij-cog, nws nyuaj dua los ua ib qho kev pab. Thiab, los ntawm txoj kev, hauv daim ntawv tshwj xeeb no, kuv tau hais meej meej tias qhov no tsis yog daim ntawv tshawb fawb hauv qhov kev nkag siab niaj hnub ntawm lo lus. Nws yog ntau ntawm daim ntawv qhia txog qhov twg qhov no yuav tsum mus. Thiab muaj ob peb lub tswv yim muaj uas tej zaum yuav tshiab, tab sis feem ntau nws tsis yog. Kuv tsis thov qhov tseem ceeb ntawm feem ntau ntawm qhov kuv tau sau rau hauv daim ntawv, qhov tseem ceeb.
Kev kawm txhawb zog kuj yuav tsis txaus, LeCun tuav. Cov kws tshawb fawb xws li David Silver ntawm DeepMind, uas tau tsim lub AlphaZero program uas tau paub Chess, Shogi thiab Go, yog "raws li kev nqis tes ua," saib LeCun, tab sis "feem ntau ntawm kev kawm peb ua, peb tsis ua los ntawm kev ua tiag tiag. kev nqis tes ua, peb ua nws los ntawm kev soj ntsuam. "
ZDNet: Thiab qhov ntawd yog qhov chaw zoo pib, vim kuv xav paub vim li cas koj thiaj li taug txoj kev no tam sim no? Koj xav li cas txog qhov no? Vim li cas koj thiaj xav sau qhov no?
YL: Zoo, yog li ntawd, kuv tau xav txog qhov no ntev heev, txog txoj hauv kev rau tib neeg-theem lossis tsiaj-theem-hom kev txawj ntse lossis kev kawm thiab kev muaj peev xwm. Thiab, hauv kuv cov lus hais kuv tau hais lus zoo txog qhov no tag nrho uas ob qho tib si saib xyuas kev kawm thiab kev txhawb nqa kev kawm tsis txaus los ua raws li hom kev kawm uas peb saib hauv cov tsiaj thiab tib neeg. Kuv tau ua qhov no rau ib yam dab tsi zoo li xya lossis yim xyoo. Yog li, tsis ntev los no. Kuv muaj cov lus tseem ceeb ntawm NeurIPS ntau xyoo dhau los uas kuv tau ua qhov ntawd, qhov tseem ceeb, thiab ntau yam kev hais lus, muaj cov ntaub ntawv kaw tseg. Tam sim no, vim li cas thiaj sau ib daim ntawv tam sim no? Kuv tau los txog - [Google Brain tus kws tshawb fawb] Geoff Hinton tau ua ib yam dab tsi zoo sib xws - Kuv txhais tau tias, qhov tseeb, nws ntau dua kuv, peb pom lub sijhawm khiav tawm. Peb tsis yog hluas.
ZDNet: Rau caum yog tus tshiab tsib caug.
YL: Qhov ntawd yog qhov tseeb, tab sis lub ntsiab lus yog, peb pom ntau qhov kev thov hais tias peb yuav tsum ua li cas txhawm rau thawb mus rau tib neeg qib ntawm AI. Thiab muaj cov tswv yim uas kuv xav tias yog misdirected. Yog li, ib lub tswv yim yog, Huag, peb yuav tsum cia li ntxiv cov cim kev xav rau saum cov neural nets. Thiab kuv tsis paub yuav ua li cas rau qhov no. Yog li, tej zaum qhov kuv tau piav qhia hauv daim ntawv yuav yog ib txoj hauv kev uas yuav ua tib yam yam tsis muaj kev qhia meej meej. Qhov no yog hom kev lig kev cai Gary Marcuses ntawm lub ntiaj teb. Gary Marcus tsis yog tus neeg AI, los ntawm txoj kev, nws yog tus kws kho mob hlwb. Nws yeej tsis tau pab dab tsi rau AI. Nws tau ua haujlwm zoo tiag tiag hauv kev sim kev puas siab puas ntsws tab sis nws tsis tau sau ib daim ntawv tshuaj xyuas cov phooj ywg ntawm AI. Yog li, muaj cov neeg ntawd.
Muaj [DeepMind tus kws tshawb fawb txog lub hauv paus ntsiab lus] David Silvers ntawm lub ntiaj teb uas hais tias, koj paub, nqi zog txaus, qhov tseem ceeb, nws yog txhua yam hais txog kev txhawb nqa kev kawm, peb tsuas yog yuav tsum ua kom nws muaj txiaj ntsig me ntsis, ua li cas? Thiab, kuv xav tias lawv tsis yog yuam kev, tab sis kuv xav tias cov kauj ruam tsim nyog los ua kom muaj kev txhawb nqa kev kawm tau zoo dua, qhov tseem ceeb, yuav tshem tawm qhov kev txhawb nqa kev kawm kom xaiv cov txiv ntoo ntawm lub ncuav mog qab zib. Thiab qhov tseem ceeb uas ploj lawm yog kawm seb lub ntiaj teb ua haujlwm li cas, feem ntau yog los ntawm kev soj ntsuam yam tsis muaj kev nqis tes ua. Kev txhawb zog kev kawm yog kev nqis tes ua heev, koj kawm tej yam hauv ntiaj teb los ntawm kev ua thiab pom cov txiaj ntsig.
ZDNet: Thiab nws yog nqi zog tsom.
YL: Nws yog nqi zog-focused, thiab nws yog action-focused thiab. Yog li ntawd, koj yuav tsum tau ua nyob rau hauv lub ntiaj teb no kom muaj peev xwm kawm tau ib yam dab tsi txog lub ntiaj teb no. Thiab qhov kev thov tseem ceeb kuv ua hauv daim ntawv hais txog kev saib xyuas tus kheej yog, feem ntau ntawm kev kawm peb ua, peb tsis ua nws los ntawm kev ua tiag tiag, peb ua los ntawm kev soj ntsuam. Thiab nws yog unorthodox heev, ob qho tib si rau kev txhawb nqa kev kawm tib neeg, tshwj xeeb tshaj yog, tab sis kuj rau ntau tus kws tshawb fawb txog kev puas siab puas ntsws thiab cov kws paub txog kev txawj ntse uas xav tias, koj paub, kev nqis tes ua - Kuv tsis hais tias kev nqis tes ua tsis tseem ceeb, nws is qhov tseem ceeb. Tab sis kuv xav tias feem ntau ntawm qhov peb kawm feem ntau yog hais txog cov qauv ntawm lub ntiaj teb, thiab koom nrog, ntawm chav kawm, kev sib cuam tshuam thiab kev ua thiab kev ua si, thiab tej yam zoo li ntawd, tab sis ntau ntawm nws yog kev soj ntsuam.
ZDNet: Koj tseem yuav tswj kom zuam tawm cov neeg Transformer, cov lus-thawj cov neeg, tib lub sijhawm. Koj yuav ua li cas thiaj tsim tau qhov no yam tsis muaj lus ua ntej? Koj tuaj yeem tswj kom zuam tawm ntau tus neeg.
YL: Yog, kuv tau siv rau qhov ntawd. Yog li, yog, muaj cov lus-thawj cov neeg, uas hais, koj paub, kev txawj ntse yog hais txog lus, lub substrate ntawm kev txawj ntse yog lus, blah, blah, blah. Tab sis qhov ntawd, zoo-ntawm, tso tseg kev txawj ntse tsiaj. Koj paub, peb tsis mus rau qhov chaw uas peb cov tshuab ntse muaj kev nkag siab ntau npaum li miv. Yog li, vim li cas peb tsis pib muaj? Nws yog dab tsi uas tso cai rau miv los ntes lub ntiaj teb ib puag ncig, ua tej yam ntse zoo nkauj, thiab npaj thiab khoom zoo li ntawd, thiab dev txawm zoo dua?
Tom qab ntawd muaj tag nrho cov neeg hais tias, Auj, kev txawj ntse yog kev sib raug zoo, puas yog? Peb txawj ntse vim peb sib tham thiab sib pauv ntaub ntawv, thiab blah, blah, blah. Muaj txhua hom tsiaj uas tsis muaj kev sib raug zoo uas tsis tau ntsib lawv niam lawv txiv uas ntse heev, zoo li octopus lossis orangutans.Kuv txhais tau tias, lawv [orangutans] yeej tau kawm los ntawm lawv niam, tab sis lawv tsis yog tsiaj txhu.
Tab sis lwm pawg ntawm cov neeg uas kuv yuav zuam tawm yog cov neeg uas hais tias scaling yog txaus. Yog li, hauv paus, peb tsuas yog siv cov Transformers loj heev, peb cob qhia lawv ntawm cov ntaub ntawv ntau yam uas suav nrog, koj paub, video, ntawv nyeem, blah, blah, blah. Peb, zoo-ntawm, petrifytxhua yam, thiab tokenize txhua yam, thiab ces cob qhia giganticcov qauv los ua qhov kev kwv yees tsis sib xws, hauv paus, thiab qee qhov AI yuav tawm ntawm qhov no. Lawv tsis yog lawm, nyob rau hauv lub siab hais tias tej zaum yuav yog ib feem ntawm lub neej yav tom ntej ntse system. Tab sis kuv xav tias nws tsis muaj cov khoom tseem ceeb.
Muaj lwm pawg neeg kuv yuav zuam tawm nrog daim ntawv no. Thiab nws yog cov probabilists, kev ntseeg probabilists. Yog li, cov neeg uas xav tias qhov kev xav yuav tshwm sim yog tib lub hauv paus uas koj tuaj yeem siv los piav qhia txog kev kawm tshuab. Thiab raws li kuv tau sim piav qhia hauv daim ntawv, nws yog qhov pib ntau dhau los thov kom tus qauv ntiaj teb ua rau muaj kev cuam tshuam tag nrho. Peb tsis paub yuav ua li cas. Muaj qhov kev suav tsis txheeb. Yog li kuv thov kom tso tag nrho lub tswv yim no. Thiab ntawm chav kawm, koj paub, qhov no yog ib qho tseem ceeb ntawm tsis tsuas yog kev kawm tshuab, tab sis tag nrho cov txheeb cais, uas tau lees tias yog ib txwm formalism rau kev kawm tshuab.
Lwm qhov -
ZDNet: Koj nyob ntawm roll…
YL: - yog qhov hu ua generative qauv. Yog li, lub tswv yim uas koj tuaj yeem kawm los kwv yees, thiab tej zaum koj tuaj yeem kawm ntau yam hauv ntiaj teb los ntawm kev twv ua ntej. Yog li, kuv muab ib daim vis dis aus rau koj thiab kuv hais kom lub kaw lus twv seb yuav muaj dab tsi tshwm sim tom ntej hauv video. Thiab kuv tuaj yeem hais kom koj kwv yees qhov tseeb video thav duab nrog txhua cov ntsiab lus. Tab sis qhov kuv sib cav txog hauv daim ntawv yog qhov ua tau ntau dhau los nug thiab nyuaj heev. Thiab qhov no yog ib yam uas kuv tau hloov kuv lub siab txog. Txog txog li ob xyoos dhau los, kuv tau los ua tus tawm tswv yim txog qhov kuv hu ua latent variable generative qauv, cov qauv uas kwv yees tias yuav muaj dab tsi tshwm sim tom ntej lossis cov ntaub ntawv uas ploj lawm, tejzaum nws nrog kev pab los ntawm qhov sib txawv latent, yog tias qhov kev twv ua tsis tau. kev txiav txim siab. Thiab kuv tau tso tseg rau qhov no. Thiab yog vim li cas kuv tau muab tso rau qhov no yog ua raws li cov txiaj ntsig pom, qhov twg tib neeg tau sim ua ntawv thov, xaiv, kwv yees lossis rov tsim kho raws li kev cob qhia ntawm hom uas siv hauv BERTthiab cov qauv lus loj, lawv tau sim siv qhov no rau cov duab, thiab nws tau ua tsis tiav. Thiab yog vim li cas nws yog qhov ua tsis tiav yog, dua, vim yog qhov txwv ntawm cov qauv probabilistic uas nws yooj yim los twv seb cov tokens zoo li cov lus vim peb tuaj yeem suav qhov muaj feem cuam tshuam rau txhua lo lus hauv phau ntawv txhais lus. Qhov ntawd yog qhov yooj yim. Tab sis yog tias peb thov kom lub kaw lus tsim qhov muaj feem cuam tshuam rau txhua qhov ua tau video thav duab, peb tsis muaj lub tswv yim yuav ua li cas parameterize nws, lossis peb muaj qee lub tswv yim yuav ua li cas parameterize nws, tab sis peb tsis paub yuav ua li cas normalize nws. Nws tsoo ib qho teeb meem kev ua lej tsis txaus ntseeg uas peb tsis paub yuav daws li cas.
Lecun hais tias "Peb tsis mus txog qhov twg peb lub tshuab ntse muaj kev nkag siab zoo ib yam li miv," Lecun hais. “Yog li, vim li cas peb tsis pib muaj? Nws yog dab tsi uas tso cai rau miv los ntes lub ntiaj teb ib puag ncig, ua tej yam ntse zoo nkauj, thiab npaj thiab khoom zoo li ntawd, thiab dev txawm zoo dua?
Yog li, yog vim li cas kuv thiaj hais tias cia peb tso tseg qhov kev xav tau los yog lub hauv paus rau tej yam zoo li ntawd, qhov tsis muaj zog, cov qauv siv zog. Kuv tau tawm tswv yim rau qhov no, thiab, tau ntau xyoo lawm, yog li qhov no tsis yog qhov tsis ntev los no. Tab sis tib lub sijhawm, tso tseg lub tswv yim ntawm cov qauv tsim vim muaj ntau yam hauv ntiaj teb uas tsis to taub thiab tsis tuaj yeem kwv yees. Yog koj yog ib tug engineer, koj hu nws nrov nrov. Yog koj yog physicist, koj hu nws kub. Thiab yog tias koj yog ib tus neeg kawm tshuab, koj hu nws, koj paub, cov ntsiab lus tsis cuam tshuam lossis dab tsi.
Yog li, qhov piv txwv kuv tau siv hauv daim ntawv, lossis kuv tau siv hauv kev hais lus, yog, koj xav tau lub ntiaj teb kev twv ua ntej uas yuav pab tau hauv tsheb tsav tus kheej, txoj cai? Nws xav kom muaj peev xwm twv tau, ua ntej, txoj hauv kev ntawm tag nrho lwm lub tsheb, yuav ua li cas rau lwm yam khoom uas yuav txav mus los, cov neeg taug kev, tsheb kauj vab, menyuam yaus khiav tom qab ncaws pob, yam zoo li ntawd. Yog li, txhua yam ntawm lub ntiaj teb. Tab sis ciam teb ntawm txoj kev, tej zaum yuav muaj ntoo, thiab muaj cua hnub no, yog li cov nplooj txav mus rau hauv cua, thiab tom qab cov ntoo muaj ib tug pas dej, thiab muaj ripples nyob rau hauv lub pas dej. Thiab cov no yog, qhov tseem ceeb, ntau yam tshwm sim tsis tau pom dua. Thiab, koj tsis xav kom koj tus qauv siv nyiaj ntau ntawm cov peev txheej kwv yees cov khoom uas nyuaj rau kwv yees thiab tsis cuam tshuam. Yog li ntawd yog vim li cas kuv thiaj li tawm tswv yim rau kev sib koom embedding architecture, cov khoom uas qhov sib txawv koj tab tom sim ua qauv, koj tsis tau sim twv nws, koj tab tom sim ua qauv, tab sis nws khiav los ntawm tus encoder, thiab uas encoder tuaj yeem tshem tawm ntau cov ntsiab lus hais txog cov tswv yim uas tsis cuam tshuam lossis nyuaj dhau - qhov tseem ceeb, sib npaug rau lub suab nrov.
ZDNet: Peb tau sib tham ua ntej xyoo no lub zog-raws li qauv, JEPA thiab H-JEPA. Kuv qhov kev nkag siab, yog tias kuv nkag siab koj raug, koj puas pom lub ntsiab lus ntawm lub zog qis qhov twg ob qhov kev kwv yees ntawm X thiab Y embeddings zoo tshaj plaws, uas txhais tau hais tias yog muaj ib tug nquab nyob rau hauv ib tsob ntoo nyob rau hauv ib tug, thiab muaj ib yam dab tsi nyob rau hauv lub keeb kwm ntawm ib qho chaw, cov no yuav tsis yog cov ntsiab lus tseem ceeb uas ua rau cov embeddings nyob ze rau ib leeg.
YL: Txoj cai. Yog li, JEPA architecture yeej sim nrhiav kev sib pauv, kev sib haum xeeb, ntawm kev rho tawm cov sawv cev uas yog cov ntaub ntawv ntau tshaj plaws txog cov khoom siv tab sis kuj tuaj yeem kwv yees los ntawm ib leeg nrog qee theem ntawm qhov raug lossis kev ntseeg tau. Nws nrhiav kev sib pauv. Yog li, yog tias nws muaj kev xaiv ntawm kev siv nyiaj ntau ntawm cov peev txheej nrog rau cov ntsiab lus ntawm cov lus tsa suab ntawm cov nplooj, thiab tom qab ntawd ua qauv ntawm cov dynamics uas yuav txiav txim siab seb cov nplooj tsiv li cas ib pliag los ntawm tam sim no, los yog cia li tso rau hauv pem teb los ntawm tsuas yog pib khiav Y qhov sib txawv los ntawm kev kwv yees uas tshem tawm tag nrho cov ntsiab lus, nws yuav tsuas yog tshem tawm nws vim tias nws tsuas yog nyuaj rau ua qauv thiab ntes.
ZDNet: Ib qho uas xav tsis thoob yog koj tau ua tus txhawb nqa zoo ntawm kev hais tias "Nws ua haujlwm, peb yuav txheeb xyuas tom qab qhov kev xav ntawm thermodynamics los piav qhia nws." Ntawm no koj tau coj ib txoj hauv kev los ntawm, "Kuv tsis paub yuav ua li cas peb yuav tsum daws qhov no, tab sis kuv xav muab qee cov tswv yim los xav txog nws," thiab tej zaum txawm mus txog qhov kev xav lossis kev xav, ntawm tsawg kawg. Qhov ntawd yog qhov nthuav vim tias muaj ntau tus neeg siv nyiaj ntau ntawm lub tsheb uas tuaj yeem pom cov neeg taug kev txawm tias lub tsheb puas muaj kev nkag siab zoo. Thiab kuv xav tias qee tus ntawm cov neeg ntawd yuav yog, tsis zuam tawm, tab sis lawv yuav hais tias, "Zoo, peb tsis quav ntsej yog tias nws tsis muaj kev nkag siab zoo, peb tau tsim qhov simulation, simulation yog amazing, thiab peb tab tom yuav txhim kho, peb yuav ua kom ntsuas qhov simulation. "
Thiab yog li nws yog qhov nthuav tias koj nyob hauv txoj haujlwm tam sim no hais, cia peb rov qab los xav txog qhov peb tab tom ua. Thiab kev lag luam yog hais tias peb tsuas yog mus rau scale, scale, scale, scale, vim hais tias crank yeej ua hauj lwm. Kuv txhais tau tias, lub semiconductor crank ntawm GPUs tiag tiag ua haujlwm.
YL: Muaj, zoo li, tsib nqe lus nug nyob ntawd. Yog li, kuv txhais tau tias, scaling yog tsim nyog. Kuv tsis thuam qhov tseeb tias peb yuav tsum ntsuas. Peb yuav tsum scale. Cov neural nets tau zoo dua thaum lawv loj dua. Tsis muaj lus nug peb yuav tsum ntsuas. Thiab cov uas yuav muaj qee theem ntawm kev nkag siab zoo yuav loj. Tsis muaj txoj kev nyob ib ncig ntawm qhov ntawd, kuv xav. Yog li scaling yog qhov zoo, nws yog qhov tsim nyog, tab sis tsis txaus. Qhov ntawd yog qhov kuv tab tom ua. Nws tsis yog scaling xwb. Qhov ntawd yog thawj qhov taw tes.
Qhov thib ob point, seb txoj kev xav los ua ntej thiab tej yam zoo li ntawd. Yog li, kuv xav tias muaj cov ntsiab lus uas tuaj ua ntej, koj yuav tsum tau nqis ib kauj ruam rov qab thiab hais tias, ua li cas, peb ua tus ntaiv no, tab sis peb xav mus rau lub hli thiab tsis muaj txoj hauv kev uas tus ntaiv no mus rau peb. Yog li, qhov tseem ceeb, qhov kuv sau ntawm no yog, peb yuav tsum tsim cov foob pob hluav taws. Kuv tsis tuaj yeem muab koj cov ntsiab lus ntawm yuav ua li cas peb tsim foob pob hluav taws, tab sis ntawm no yog cov hauv paus ntsiab lus. Thiab kuv tsis yog sau ib txoj kev xav rau nws los yog dab tsi, tab sis, nws yuav yog foob pob ua ntxaij, ok? Los yog qhov chaw elevator los yog dab tsi. Tej zaum peb yuav tsis muaj tag nrho cov ntsiab lus ntawm tag nrho cov tshuab. Peb tab tom sim ua kom qee yam ua haujlwm, zoo li kuv tau ua haujlwm ntawm JEPA. Kev sib koom ua ke ua haujlwm zoo heev rau kev paub txog cov duab, tab sis siv nws los cob qhia lub ntiaj teb qauv, muaj teeb meem. Peb tab tom ua haujlwm rau nws, peb cia siab tias peb yuav ua kom nws ua haujlwm soon, tab sis tej zaum peb yuav ntsib qee yam teeb meem nyob ntawd uas peb tsis tuaj yeem hla, tej zaum.
Tom qab ntawd muaj ib lub tswv yim tseem ceeb hauv daim ntawv hais txog kev xav qhov twg yog tias peb xav kom cov txheej txheem tuaj yeem npaj tau, uas koj tuaj yeem xav tias yog ib daim ntawv yooj yim ntawm kev xav, lawv yuav tsum muaj qhov sib txawv latent. Nyob rau hauv lwm yam lus, tej yam uas tsis xam los ntawm tej neural net tab sis tej yam uas yog - nws tus nqi yog inferred thiaj li yuav txo tau ib co hom phiaj muaj nuj nqi, ib co nqi muaj nuj nqi. Thiab tom qab ntawd koj tuaj yeem siv tus nqi no los tsav tus cwj pwm ntawm qhov system. Thiab qhov no tsis yog lub tswv yim tshiab, txoj cai? Qhov no yog classical, kev pom zoo tswj qhov twg lub hauv paus ntawm qhov no rov qab mus rau lig '50s, thaum ntxov '60s. Yog li ntawd, tsis txhob thov tej novelty ntawm no. Tab sis qhov kuv tab tom hais yog tias hom kev xav no yuav tsum yog ib feem ntawm qhov kev txawj ntse uas muaj peev xwm ntawm kev npaj, thiab nws tus cwj pwm tuaj yeem raug teev los yog tswj tsis tau los ntawm tus cwj pwm nyuaj, tsis yog los ntawm imitation leaning, tab sis los ntawm lub hom phiaj ua haujlwm uas tsav tus cwj pwm - tsis tsav kev kawm, tsim nyog, tab sis nws tsav tus cwj pwm. Koj paub, peb muaj qhov ntawd hauv peb lub hlwb, thiab txhua tus tsiaj muaj tus nqi tsim nyog lossis kev txhawb siab rau txhua yam. Qhov ntawd ua rau cov menyuam muaj cuaj hli xav sawv ntsug. Tus nqi ntawm kev zoo siab thaum koj sawv, lub sij hawm ntawd nyob rau hauv tus nqi muaj nuj nqi yog hardwired. Tab sis yuav ua li cas koj sawv tsis yog, qhov ntawd yog kev kawm.
"Scaling yog qhov zoo, nws yog qhov tsim nyog, tab sis tsis txaus," hais tias LeCun ntawm cov qauv lus loj xws li Transformer-based programs ntawm GPT-3 ntau yam. Transformer devotées ntseeg hais tias, "Peb tokenize txhua yam, thiab cob qhia loj heevcov qauv los ua qhov kev kwv yees tsis sib xws, thiab qee qhov AI yuav tawm ntawm qhov no ... tab sis kuv xav tias nws tsis muaj qhov tseem ceeb. "
ZDNet: Tsuas yog txhawm rau hais txog qhov ntawd, ntau lub zej zog kev kawm tob zoo li zoo mus tom ntej nrog qee yam uas tsis muaj kev nkag siab zoo. Nws zoo li koj tab tom ua qhov kev sib cav zoo nkauj ntawm no tias qee lub sijhawm nws dhau los ua qhov tsis txaus ntseeg. Qee tus neeg hais tias peb tsis xav tau lub tsheb uas muaj kev ywj pheej nrog kev nkag siab zoo vim tias kev ntsuas yuav ua rau nws. Nws zoo li koj hais tias nws tsis ua li cas cia li taug txoj kev ntawd?
YL: Koj paub, kuv xav tias nws muaj peev xwm tag nrho uas peb yuav muaj qib tsib lub tsheb uas tsis muaj kev nkag siab zoo. Tab sis qhov teeb meem nrog rau qhov no mus kom ze, qhov no yuav yog ib ntus, vim hais tias koj yuav tsum tau engineer lub ntuj raug txim tawm ntawm nws. Yog li, koj paub, daim ntawv qhia tag nrho lub ntiaj teb, nyuaj-xaim txhua yam ntawm cov kev coj cwj pwm tshwj xeeb, sau cov ntaub ntawv txaus uas koj muaj tag nrho cov xwm txheej txawv txawv uas koj tuaj yeem ntsib ntawm txoj kev, blah, blah, blah. Thiab kuv qhov kev kwv yees yog tias nrog kev nqis peev txaus thiab lub sijhawm, koj tuaj yeem tsim lub ntuj raug txim tawm ntawm nws. Tab sis thaum kawg, yuav muaj kev txaus siab dua thiab muaj peev xwm daws tau zoo dua uas suav nrog cov tshuab ua haujlwm zoo dua ntawm kev nkag siab txog kev ua haujlwm hauv ntiaj teb, thiab muaj, koj paub, qee theem ntawm qhov peb yuav hu ua kev nkag siab zoo. Nws tsis tas yuav yog tib neeg-theem kev nkag siab zoo, tab sis qee yam kev paub uas lub kaw lus tuaj yeem tau los ntawm kev saib, tab sis tsis saib ib tus neeg tsav tsheb, tsuas yog saib cov khoom txav mus los thiab nkag siab ntau txog lub ntiaj teb, tsim lub hauv paus ntawm keeb kwm yav dhau los. kev paub txog yuav ua li cas lub ntiaj teb ua haujlwm, nyob rau sab saum toj uas koj tuaj yeem kawm tsav tsheb.
Cia kuv ua piv txwv keeb kwm ntawm qhov no. Classical computer vision yog nyob ntawm ntau hardwired, engineered modules, nyob rau sab saum toj ntawm uas koj yuav muaj, zoo-ntawm, ib tug nyias txheej ntawm kev kawm. Yog li, cov khoom uas raug ntaus los ntawm AlexNet nyob rau hauv 2012, muaj ib tug thawj theem, zoo-ntawm, handcrafted feature extractions, zoo li SIFTs [Scale-Invariant Feature Transform (SIFT), ib tug classic tsis pom kev txheej txheem los txheeb xyuas cov khoom tseem ceeb hauv ib daim duab] thiab HOG [Histogram ntawm Oriented Gradints, lwm cov txheej txheem classic] thiab ntau yam. Thiab tom qab ntawd cov txheej txheem thib ob ntawm, txheeb-ntawm, nruab nrab-theem nta raws li cov ntsiab lus tshwj xeeb thiab txawm li cas los xij, thiab qee yam ntawm txoj kev tsis saib xyuas. Thiab tom qab ntawd nyob rau sab saum toj ntawm qhov no, koj muab lub tshuab vector txhawb nqa, lossis lwm yam kev faib tawm yooj yim. Thiab qhov ntawd yog, zoo-ntawm, tus qauv raj xa dej los ntawm nruab nrab-2000s mus rau 2012. Thiab qhov ntawd tau hloov los ntawm qhov kawg-rau-kawg convolutional nets, qhov twg koj tsis hardwire ib qho ntawm no, koj tsuas yog muaj ntau cov ntaub ntawv, thiab koj cob qhia txog qhov kawg mus rau qhov kawg, uas yog txoj hauv kev uas kuv tau tawm tswv yim rau lub sijhawm ntev, tab sis koj paub, txog thaum ntawd, tsis yog qhov ua tau zoo rau cov teeb meem loj.
Muaj ib zaj dab neeg zoo sib xws hauv kev paub txog kev hais lus nyob qhov twg, dua, muaj ntau cov ncauj lus kom ntxaws engineering rau koj ua ntej cov ntaub ntawv li cas, koj rho tawm qhov loj-scale cepstrum [tus hloov pauv ntawm Fast Fourier Transform rau cov teeb liab ua], thiab tom qab ntawd Koj muaj Hidden Markov Models, nrog rau kev xaiv, pre-set architecture, blah, blah, blah, nrog Mixture ntawm Gaussians. Thiab yog li ntawd, nws yog ib qho me ntsis ntawm tib lub architecture raws li qhov pom kev uas koj tau xuas tes ua pem hauv ntej-kawg, thiab tom qab ntawd ib qho me ntsis tsis saib xyuas, kev cob qhia, txheej nruab nrab, thiab tom qab ntawd ib txheej saib xyuas rau saum. Thiab tam sim no qhov ntawd tau, qhov tseem ceeb, tshem tawm los ntawm qhov kawg-rau-kawg neural nets. Yog li kuv pom ib yam dab tsi zoo sib xws ntawm kev sim kawm txhua yam, tab sis koj yuav tsum muaj txoj cai ua ntej, txoj cai architecture, txoj cai qauv.
Cov neeg tsav tsheb tus kheej, kev pib ua lag luam xws li Waymo thiab Wayve, tau "zoo siab me ntsis," nws hais tias, los ntawm kev xav tias lawv tuaj yeem "pov cov ntaub ntawv ntawm nws, thiab koj tuaj yeem kawm tau ntau yam." Lub tsheb tsav tus kheej ntawm Qib 5 ntawm ADAS muaj peev xwm ua tau, "Tab sis koj yuav tsum tau tsim lub ntuj raug txim tawm ntawm nws" thiab yuav "nkim" zoo li cov qauv hauv computer pom thaum ntxov.
ZDNet: Qhov koj tab tom hais yog, qee tus neeg yuav sim ua kom tiav qhov tsis ua haujlwm tam sim no nrog kev kawm tob rau kev siv tau, hais, hauv kev lag luam, thiab lawv yuav pib tsim ib yam dab tsi uas yog qhov uas tsis siv hauv computer tsis pom kev?
YL: Txoj cai. Thiab nws yog ib feem yog vim li cas cov neeg ua haujlwm ntawm kev tsav tsheb tsis muaj zog tau ua rau muaj kev cia siab me ntsis dhau ob peb xyoos dhau los, vim tias, koj paub, koj muaj cov no, cov khoom lag luam zoo li cov nets thiab Transformers, uas koj tuaj yeem pov cov ntaub ntawv ntawm nws. , thiab nws tuaj yeem kawm tau ntau npaum li cas. Yog li, koj hais, Okay, Kuv muaj kev daws teeb meem rau qhov teeb meem ntawd. Thawj qhov koj ua yog koj tsim lub demo qhov twg lub tsheb tsav nws tus kheej li ob peb feeb yam tsis ua mob rau leej twg. Thiab tom qab ntawd koj paub tias muaj ntau lub kaum rooj plaub, thiab koj sim npaj qhov nkhaus ntawm qhov zoo npaum li cas kuv tau txais thaum kuv muab ob npaug rau kev cob qhia, thiab koj paub tias koj yuav tsis mus rau qhov ntawd vim tias muaj txhua yam ntawm kaum rooj plaub. . Thiab koj yuav tsum muaj lub tsheb uas yuav ua rau muaj kev sib tsoo tuag tsawg dua txhua 200 lab kilometers, puas yog? Yog li, koj ua li cas? Zoo, koj taug kev hauv ob txoj kev.
Thawj qhov kev taw qhia yog, yuav ua li cas thiaj li txo tau cov ntaub ntawv uas tsim nyog rau kuv qhov system kawm? Thiab qhov ntawd yog qhov uas kev saib xyuas tus kheej tuaj rau hauv. Yog li, ntau lub tsheb tsav tsheb tus kheej tau txaus siab heev rau kev saib xyuas tus kheej vim tias qhov ntawd yog ib txoj hauv kev tseem siv cov ntaub ntawv saib xyuas loj heev rau kev kawm ua yeeb yam, tab sis tau txais kev ua tau zoo dua los ntawm pre-kev cob qhia, qhov tseem ceeb. Thiab nws tseem tsis tau panned heev, tab sis nws yuav. Thiab tom qab ntawd muaj lwm txoj kev xaiv, uas feem ntau ntawm cov tuam txhab uas muaj kev tshaj lij ntawm lub sijhawm no tau txais yuav, uas yog, tsis ua li cas, peb tuaj yeem ua qhov kawg-rau-kawg kev cob qhia, tab sis muaj ntau lub ces kaum uas peb tuaj yeem ua tau. t kov, yog li peb yuav tsuas yog engineer systems uas yuav saib xyuas cov ces kaum rooj plaub, thiab, qhov tseem ceeb, kho lawv li tshwj xeeb kis, thiab hardwire tswj, thiab ces hardwire ib tug ntau ntawm cov kev coj cwj pwm los lis cov xwm txheej tshwj xeeb. Thiab yog tias koj muaj ib pab neeg ua haujlwm loj txaus, koj tuaj yeem rub nws tawm. Tab sis nws yuav siv sij hawm ntev, thiab thaum kawg, nws tseem yuav nkig me ntsis, tej zaum yuav txhim khu kev qha txaus uas koj tuaj yeem siv tau, tab sis nrog qee theem ntawm brittleness, uas, nrog rau kev kawm ntau dua uas yuav tshwm sim hauv yav tom ntej, tsheb yuav tsis muaj vim nws yuav muaj qee qhov kev nkag siab zoo thiab kev nkag siab txog kev ua haujlwm hauv ntiaj teb.
Nyob rau hauv lub sij hawm luv luv, qhov, yam-ntawm, engineered mus kom ze yuav yeej - nws twb yeej. Qhov ntawd yog Waymo thiab Cruise ntawm lub ntiaj teb thiab Wayvethiab txawm li cas los xij, qhov ntawd yog qhov lawv ua. Tom qab ntawd muaj txoj hauv kev los saib xyuas tus kheej, uas tej zaum yuav pab tau txoj hauv kev tsim kho kom ua tiav. Tab sis tom qab ntawd, nyob rau hauv lub sij hawm ntev, uas tej zaum yuav ntev dhau rau cov tuam txhab yuav tsum tau tos, tej zaum yuav, zoo-ntawm, ib tug ntau integrated autonomous ntse tsav tsheb.
ZDNet: Peb hais dhau lub peev txheej ntawm cov tub ua lag luam feem ntau.
YL: Yog lawm. Yog li, lo lus nug yog, cov neeg puas yuav poob siab lossis tsis muaj nyiaj ua ntej qhov kev ua tau zoo mus txog qib uas xav tau.
ZDNet: Puas muaj ib yam dab tsi nthuav qhia txog vim li cas koj xaiv qee lub ntsiab lus koj xaiv hauv tus qauv? Vim koj hais Kenneth Craik [1943,Qhov Kev piav qhia], thiab koj hais txog Bryson thiab Ho [1969, Siv kev tswj kom zoo], thiab kuv xav paub vim li cas koj pib nrog cov kev cuam tshuam no, yog tias koj ntseeg tshwj xeeb tshaj yog tias cov neeg no tau ntsia nws kom deb li deb li lawv tau ua. Vim li cas koj thiaj pib muaj?
YL: Zoo, kuv tsis xav tias, yeej, lawv muaj tag nrho cov ntsiab lus nailed. Yog li, Bryson thiab Ho, qhov no yog ib phau ntawv uas kuv nyeem rov qab rau xyoo 1987 thaum kuv yog postdoc nrog Geoffrey Hinton hauv Toronto. Tab sis kuv paub txog txoj haujlwm no ua ntej thaum kuv tab tom sau kuv PhD, thiab ua rau muaj kev sib txuas ntawm kev tswj hwm thiab backprop, qhov tseem ceeb. Yog tias koj xav ua tiag tiag, koj paub, lwm tus Schmidhuber, koj yuav hais tias cov neeg tsim khoom ntawm backprop tiag tiag yog qhov zoo tshaj plaws tswj theorists Henry J. Kelley, Arthur Bryson, thiab tej zaum txawm tias Lev Pontryagin, uas yog Lavxias teb sab theorist ntawm kev pom zoo rov qab. nyob rau hauv lub lig '50s.
Yog li, lawv xam pom nws, thiab qhov tseeb, koj tuaj yeem pom lub hauv paus ntawm qhov no, cov lej hauv qab ntawd, yog Lagrangian mechanics. Yog li koj tuaj yeem rov qab mus rau Euler thiab Lagrange, qhov tseeb, thiab hom kev nrhiav ib qho ntawm qhov no hauv lawv lub ntsiab lus ntawm Lagrangian classical mechanics, tiag tiag. Yog li, nyob rau hauv cov ntsiab lus ntawm kev tswj kom zoo, dab tsi hais mav xav nyob rau hauv yog pib xam lub foob pob ua ntxaij trajectories. Koj paub, qhov no yog lub hnub nyoog thaum ntxov. Thiab yog tias koj muaj tus qauv ntawm lub foob pob hluav taws, nws qhia rau koj ntawm no yog lub xeev ntawm lub foob pob hluav taws thaum lub sijhawm t, thiab ntawm no yog qhov kuv yuav ua, yog li, thrust thiab actuators ntawm ntau yam, ntawm no yog lub xeev ntawm lub foob pob hluav taws thaum lub sijhawm t + 1.
ZDNet: Ib lub xeev-ua qauv qauv, tus qauv nqi.
YL: Yog lawm, lub hauv paus ntawm kev tswj. Yog li, tam sim no koj tuaj yeem simulate qhov kev tua ntawm koj lub foob pob hluav taws los ntawm kev xav txog ib ntu ntawm cov lus txib, thiab tom qab ntawd koj muaj qee tus nqi ua haujlwm, uas yog qhov kev ncua deb ntawm lub foob pob hluav taws mus rau nws lub hom phiaj, chaw nres tsheb lossis txawm tias nws yog. Thiab tom qab ntawd los ntawm qee yam ntawm qhov gradient qhovntsej thiaj tsis mob, koj tuaj yeem txiav txim siab, yuav ua li cas kuv tuaj yeem hloov kho kuv cov txheej txheem ntawm kev ua kom kuv lub foob pob hluav taws tau ze li sai tau rau lub hom phiaj. Thiab qhov ntawd yuav tsum tau los ntawm rov qab-propagating signals rov qab rau lub sijhawm. Thiab qhov ntawd yog back-propagation, gradient back-propagation. Cov teeb liab, lawv hu ua conjugate variables hauv Lagrangian mechanics, tab sis qhov tseeb, lawv yog gradients. Yog li, lawv tau tsim cov khoom siv rov qab, tab sis lawv tsis paub tias txoj cai no tuaj yeem siv los cob qhia ntau theem uas tuaj yeem ua tus qauv lees paub lossis qee yam zoo li ntawd. Qhov no tsis yog tiag tiag kom txog rau thaum tej zaum lig '70s, thaum ntxov '80s, thiab tom qab ntawd tsis tau siv tiag tiag thiab ua haujlwm kom txog rau thaum nruab nrab-80s. Okay, yog li ntawd, qhov no yog qhov backprop tiag tiag, zoo-ntawm, tau tawm vim tias tib neeg tau pom ntawm no yog ob peb kab ntawm cov cai uas koj tuaj yeem cob qhia cov neural net, kawg mus rau qhov kawg, multilayer. Thiab qhov ntawd txhawb nqa cov kev txwv ntawm Perceptron. Thiab, yog, muaj kev sib txuas nrog kev tswj kom zoo, tab sis qhov ntawd tsis yog.
ZDNet: Yog li, qhov ntawd yog txoj kev ntev ntawm kev hais tias cov kev cuam tshuam uas koj tau pib nrog tau rov qab mus rau backprop, thiab qhov tseem ceeb yog qhov pib rau koj?
YL: Yog lawm, tab sis kuv xav tias tib neeg tsis nco qab me ntsis txog, muaj kev ua haujlwm me ntsis ntawm qhov no, koj paub, rov qab rau xyoo 90s, lossis txawm tias xyoo 80s, suav nrog cov neeg zoo li Michael Jordan [MIT Dept. of Brain thiab Cognitive Sciences] thiab cov neeg nyiam cov uas tsis ua neural nets ntxiv lawm, tab sis lub tswv yim tias koj tuaj yeem siv neural nets rau kev tswj, thiab koj tuaj yeem siv cov tswv yim classical ntawm kev tswj kom zoo. Yog li, tej yam zoo li qhov hu ua qauv-kev tswj hwm kev twv twv, tam sim no hu ua qauv-kev tswj hwm kev twv ua ntej, lub tswv yim no uas koj tuaj yeem simulate lossis xav txog qhov tshwm sim ntawm ib ntu ntawm kev ua yog tias koj muaj tus qauv zoo ntawm qhov koj tab tom sim tswj. Thiab ib puag ncig nws nyob rau hauv. Thiab tom qab ntawd los ntawm qhov gradient qhovntsej thiaj tsis mob, qhov tseem ceeb - qhov no tsis yog kev kawm, qhov no yog kev xav - koj tuaj yeem txiav txim siab seb qhov twg yog qhov zoo tshaj plaws ntawm kev ua uas yuav txo kuv lub hom phiaj. Yog li, kev siv tus nqi ua haujlwm nrog qhov sib txawv latent rau inference yog, kuv xav, ib yam dab tsi uas tam sim no cov qoob loo ntawm loj-scale neural nets tsis nco qab txog. Tab sis nws yog ib qho tseem ceeb heev ntawm kev kawm tshuab tau ntev. Yog li, txhua tus Bayesian Net lossis graphical model lossis probabilistic graphical qauv siv hom kev xav no. Koj muaj ib tug qauv uas captures lub dependencies ntawm ib pawg ntawm cov variables, koj tau qhia tus nqi ntawm ib co ntawm cov variables, thiab ces koj yuav tsum infer lub feem ntau yuav muaj nuj nqis ntawm tus so ntawm cov variables. Qhov ntawd yog lub hauv paus ntsiab lus ntawm inference nyob rau hauv graphical qauv thiab Bayesian Nets, thiab tej yam zoo li ntawd. Thiab kuv xav tias qhov tseem ceeb yog qhov kev xav yuav tsum yog hais txog, kev xav thiab kev npaj.
ZDNet: Koj yog lub txee dai khaub ncaws Bayesian.
YL: Kuv yog ib tug neeg uas tsis yog Bayesian. Kuv ua qhov tso dag ua ntej. Kuv yeej nyob ntawm NeurIPS ob peb xyoos dhau los, Kuv xav tias nws yog xyoo 2018 lossis 2019, thiab kuv tau ntes tau hauv video los ntawm Bayesian uas nug kuv tias kuv yog Bayesian, thiab kuv hais tias, Yog, Kuv yog Bayesian, tab sis kuv Kuv yog ib tug uas tsis yog-probabilistic Bayesian, xaiv-ntawm, lub zog-based Bayesian, yog tias koj xav tau.
ZDNet: Uas twv yuav raug hu zoo li ib yam dab tsi los ntawm Star Trek. Koj tau hais nyob rau hauv qhov kawg ntawm daim ntawv no, nws yuav siv sij hawm ntau xyoo ntawm kev ua haujlwm nyuaj heev kom paub qhov koj xav li cas. Qhia rau kuv txog qee yam ntawm cov haujlwm ntawd tam sim no muaj.
YL: Yog li, kuv piav qhia koj li cas cob qhia thiab tsim JEPA hauv daim ntawv. Thiab qhov kev ntsuas uas kuv tab tom tawm tswv yim yog muaj qee txoj hauv kev kom ua tiav cov ntaub ntawv cov ntsiab lus uas cov sawv cev uas tau muab rho tawm muaj txog cov tswv yim. Thiab ces qhov thib ob yog txo qhov kev twv ua ntej yuam kev. Thiab yog tias koj muaj qhov sib txawv latent nyob rau hauv tus kwv yees uas tso cai rau tus kwv yees tsis yog qhov kev txiav txim siab, koj yuav tsum ua raws li qhov kev hloov pauv latent no los ntawm kev txo qis nws cov ntsiab lus. Yog li, koj muaj ob qhov teeb meem tam sim no, uas yog li cas koj ua kom cov ntsiab lus ntawm cov zis ntawm qee qhov neural net, thiab lwm qhov yog li cas koj txo cov ntsiab lus ntawm qee qhov txawv txav? Thiab yog tias koj tsis ua ob yam no, lub kaw lus yuav vau. Nws yuav tsis kawm dab tsi nthuav. Nws yuav muab xoom lub zog rau txhua yam, ib yam li ntawd, uas tsis yog tus qauv zoo ntawm kev vam khom. Nws yog qhov teeb meem kev sib tsoo-tiv thaiv uas kuv hais.
Thiab kuv tab tom hais txog txhua yam uas tib neeg tau ua, tsuas muaj ob pawg ntawm txoj hauv kev los tiv thaiv kev vau. Ib qho yog txoj kev sib txawv, thiab lwm qhov yog cov txheej txheem tsis tu ncua. Yog li, lub tswv yim ntawm kev ua kom tiav cov ntsiab lus ntawm cov lus sawv cev ntawm ob qhov kev tawm tswv yim thiab txo qis cov ntaub ntawv cov ntsiab lus ntawm qhov sib txawv latent, uas belongs rau cov txheej txheem tsis tu ncua. Tab sis ntau txoj haujlwm hauv cov kev sib koom ua ke embedding architectures yog siv txoj kev sib txawv. Qhov tseeb, tej zaum lawv yog cov neeg nyiam tshaj plaws tam sim no. Yog li, lo lus nug yog raws nraim li cas koj ntsuas cov ntaub ntawv cov ntsiab lus nyob rau hauv ib txoj kev uas koj yuav optimize los yog txo? Thiab qhov ntawd yog qhov uas tej yam yuav nyuaj vim peb tsis paub yuav ua li cas ntsuas cov ntsiab lus. Peb tuaj yeem kwv yees nws, peb tuaj yeem sab saud, peb tuaj yeem ua yam zoo li ntawd. Tab sis lawv tsis ntsuas cov ntsiab lus ntawm cov ntaub ntawv, uas, qhov tseeb, rau qee qhov tsis tau txhais tau zoo.
ZDNet: Tsis yog Shannon Txoj Cai? Nws tsis yog ntaub ntawv theory? Koj tau txais qee qhov nyiaj ntawm entropy, zoo entropy thiab tsis zoo entropy, thiab qhov zoo entropy yog lub cim system uas ua haujlwm, tsis zoo entropy yog suab nrov. Tsis yog txhua yam daws tau los ntawm Shannon?
YL: Koj hais yog lawm, tab sis muaj qhov tsis zoo loj tom qab ntawd. Koj yog txoj cai hauv kev nkag siab tias yog tias koj muaj cov ntaub ntawv los ntawm koj thiab koj tuaj yeem ua qee qhov ntsuas cov ntaub ntawv rau hauv cov cim tsis sib xws, thiab tom qab ntawd koj ntsuas qhov tshwm sim ntawm txhua tus cim, ces qhov siab tshaj plaws ntawm cov ntaub ntawv nqa los ntawm cov cim no yog sum tshaj cov cim tau ntawm Pi log tau, txoj cai? Qhov twg Pi yog qhov tshwm sim ntawm lub cim kuv — uas yog Shannon entropy. [Shannon Txoj Cai yog feem ntau tsim raws li H = – ∑ pi log pi.]
Ntawm no yog qhov teeb meem, tab sis: Dab tsi yog Pi? Nws yog ib qho yooj yim thaum tus naj npawb ntawm cov cim me me thiab cov cim tau kos nws tus kheej. Thaum muaj ntau lub cim, thiab kev vam khom, nws nyuaj heev. Yog li, yog tias koj muaj ib ntu ntawm cov khoom thiab koj xav tias cov khoom tsis muaj kev ywj pheej ntawm ib leeg thiab qhov tshwm sim yog sib npaug ntawm ib qho thiab xoom lossis xijpeem, ces koj tuaj yeem ntsuas qhov entropy yooj yim, tsis muaj teeb meem. Tab sis yog tias tej yam uas tuaj rau koj yog cov vectors siab, zoo li, koj paub, cov ntaub ntawv thav duab, lossis ib yam dab tsi zoo li no, dab tsi yog Pi? Qhov kev faib tawm yog dab tsi? Ua ntej koj yuav tsum quantize qhov chaw, uas yog ib tug high-dimensional, nruam qhov chaw. Koj tsis muaj lub tswv yim yuav ua li cas quantize qhov no kom raug. Koj tuaj yeem siv k-txhais tau tias, thiab lwm yam. Qhov no yog qhov tib neeg ua thaum lawv ua video compression thiab duab compression. Tab sis nws tsuas yog qhov kwv yees xwb. Thiab tom qab ntawd koj yuav tsum ua qhov kev xav ntawm kev ywj pheej. Yog li, nws yog qhov tseeb tias hauv ib qho yeeb yaj kiab, cov thav ntawv txuas ntxiv tsis yog ywj pheej. Muaj kev vam khom, thiab tus ncej ntawd yuav nyob ntawm lwm tus ncej uas koj pom ib teev dhau los, uas yog daim duab ntawm tib yam. Yog li, koj paub, koj ntsuas tsis tau Pi. Kev ntsuas Pi, koj yuav tsum muaj lub tshuab kev kawm uas kawm los kwv yees. Thiab yog li koj rov qab mus rau qhov teeb meem dhau los. Yog li, koj tsuas tuaj yeem kwv yees qhov ntsuas ntawm cov ntaub ntawv, qhov tseem ceeb.
"Cov lus nug yog raws nraim li cas koj ntsuas cov ntaub ntawv cov ntsiab lus nyob rau hauv ib txoj kev uas koj yuav optimize los yog txo?" hais tias LeCun. "Thiab qhov ntawd yog qhov uas tej yam yuav nyuaj vim peb tsis paub yuav ntsuas cov ntsiab lus li cas." Qhov zoo tshaj plaws uas tuaj yeem ua tau kom deb li deb yog nrhiav tus neeg sawv cev uas "zoo txaus rau txoj haujlwm uas peb xav tau."
Cia kuv ua piv txwv ntxiv. Ib qho ntawm cov algorithm uas peb tau ua si nrog, thiab kuv tau tham txog hauv daim, yog qhov no hu ua VICReg, variance-invariance-covariance normalization. Nws yog nyob rau hauv ib daim ntawv cais uas tau luam tawm ntawm ICLR, thiab tau muab tso rau ntawm arXiv kwv yees li ib xyoos ua ntej, 2021. Thiab lub tswv yim muaj los ua kom cov ntaub ntawv ntau tshaj plaws. Thiab lub tswv yim tau tawm ntawm ib daim ntawv ua ntej los ntawm kuv pab pawg hu ua Barlow Twins. Koj maximize cov ntaub ntawv cov ntsiab lus ntawm ib tug vector tawm los ntawm ib tug neural net los ntawm, nyob rau hauv qhov tseeb, piv txwv li tias tsuas yog kev vam meej ntawm cov variables yog correlation, linear dependency. Yog li, yog tias koj xav tias qhov kev vam khom tsuas yog ua tau ntawm cov khub ntawm qhov sib txawv, lossis ntawm qhov sib txawv hauv koj lub cev, yog kev sib raug zoo ntawm khub ntawm cov khoom muaj nqis, uas yog qhov kwv yees tsis tshua muaj siab, ces koj tuaj yeem ua kom tiav cov ntaub ntawv tawm ntawm koj lub cev. los ntawm kev ua kom txhua qhov sib txawv tsis yog xoom variance - cia peb hais, qhov sib txawv ntawm ib qho, nws tsis muaj teeb meem nws yog dab tsi - thiab tom qab ntawd rov qab sib cuam tshuam lawv, tib txheej txheem uas hu ua whitening, nws tsis yog tshiab. Qhov teeb meem nrog qhov no yog tias koj tuaj yeem ua tau zoo heev muaj kev vam meej ntawm ob pawg ntawm kev sib txawv lossis txawm tias tsuas yog khub ntawm qhov sib txawv uas tsis yog cov kab sib txuas, thiab lawv tsis tshwm sim hauv kev sib raug zoo. Yog li, piv txwv li, yog tias koj muaj ob qhov sib txawv, thiab tag nrho cov ntsiab lus ntawm ob qhov sib txawv no nyob rau hauv qee yam ntawm kauv, muaj qhov muaj zog heev ntawm ob qhov sib txawv, txoj cai? Tab sis qhov tseeb, yog tias koj suav qhov kev sib raug zoo ntawm ob qhov sib txawv, lawv tsis muaj feem cuam tshuam. Yog li, ntawm no yog ib qho piv txwv uas cov ntsiab lus ntawm ob qhov sib txawv no yog qhov tsawg heev, nws tsuas yog ib qho kom muaj nuj nqis vim nws yog koj txoj hauj lwm hauv cov kauv. Lawv yog de-correlated, yog li koj xav tias koj muaj ntau cov ntaub ntawv tawm los ntawm ob qhov sib txawv no thaum qhov tseeb koj tsis, koj tsuas muaj, koj paub, koj tuaj yeem kwv yees ib qho ntawm qhov sib txawv ntawm lwm qhov, qhov tseem ceeb. Yog li, qhov ntawd qhia tau hais tias peb tsuas muaj kwv yees ntau txoj hauv kev los ntsuas cov ntaub ntawv cov ntsiab lus.
ZDNet: Thiab yog li ntawd yog ib qho ntawm cov khoom uas koj yuav tsum tau ua haujlwm tam sim no nrog qhov no? Qhov no yog lo lus nug loj ntawm yuav ua li cas peb paub thaum peb ua kom tiav thiab txo qis cov ntsiab lus?
YL: Los yog seb lub npe peb siv rau qhov no zoo txaus rau txoj haujlwm uas peb xav tau. Qhov tseeb, peb ua qhov no txhua lub sijhawm hauv kev kawm tshuab. Tus nqi ua haujlwm peb txo qis tsis yog qhov uas peb xav kom txo qis. Yog li, piv txwv li, koj xav ua kev faib tawm, ua li cas? Tus nqi ua haujlwm uas koj xav kom txo qis thaum koj cob qhia tus neeg faib khoom yog tus lej ntawm qhov yuam kev uas tus neeg ua haujlwm tau ua. Tab sis qhov ntawd yog qhov tsis sib txawv, tus nqi txaus ntshai uas koj tsis tuaj yeem txo qis vim tias koj paub tias koj yuav hloov qhov hnyav ntawm koj lub neural net, tsis muaj dab tsi yuav hloov mus txog thaum ib qho ntawm cov qauv ntawd tig nws qhov kev txiav txim siab, thiab tom qab ntawd dhia. nyob rau hauv qhov yuam kev, zoo los yog tsis zoo.
ZDNet: Yog li koj muaj lub npe uas yog lub hom phiaj ua haujlwm uas koj tuaj yeem hais tau tiag tiag, peb tuaj yeem txaus siab gradients ntawm qhov khoom no.
YL: Yog lawm. Yog li cov tib neeg siv qhov kev poob qhov kev sib tw no, lossis SOFTMAX, koj muaj ntau lub npe rau nws, tab sis nws yog tib yam. Thiab nws hauv paus yog ib qho kev kwv yees ntawm tus lej ntawm qhov yuam kev uas lub kaw lus ua, qhov twg smoothing yog ua los ntawm, pib, coj mus rau hauv tus account cov qhab nia uas lub kaw lus muab rau txhua pawg.
ZDNet: Puas muaj tej yam uas peb tsis tau them uas koj xav npog?
YL: Nws yog tej zaum hais txog cov ntsiab lus tseem ceeb. Kuv xav tias AI systems yuav tsum muaj peev xwm ua kom muaj laj thawj, thiab cov txheej txheem rau qhov no uas kuv tawm tswv yim yog txo qee lub hom phiaj nrog rau qee qhov sib txawv latent. Qhov ntawd tso cai rau cov txheej txheem los npaj thiab ua laj thawj. Kuv xav tias peb yuav tsum tso tseg lub ntsiab lus tseem ceeb vim tias nws yog qhov tsis txaus ntseeg thaum peb xav ua tej yam xws li kev ntes kev cia siab ntawm qhov siab, qhov sib txawv tsis tu ncua. Thiab kuv tab tom tawm tswv yim kom tso tseg cov qauv tsim tawm vim tias lub kaw lus yuav tau mob siab rau ntau yam kev pab cuam los kwv yees tej yam uas nyuaj dhau los kwv yees thiab tej zaum yuav siv nyiaj ntau dhau. Thiab qhov ntawd yog qhov zoo nkauj heev. Qhov ntawd yog cov lus tseem ceeb, yog tias koj xav tau. Thiab ces tag nrho cov architecture. Tom qab ntawd muaj cov speculations txog qhov xwm ntawm kev nco qab thiab lub luag hauj lwm ntawm tus configurator, tab sis qhov no yog tiag tiag speculation.
ZDNet: Peb mam li mus rau lwm zaus. Kuv yuav nug koj, koj ua li cas ntsuas qhov no? Tab sis kuv xav tias koj nyob nraum me ntsis ntxiv los ntawm benchmarking tam sim no?
YL: Tsis tas li ntawd nyob deb, xaiv-ntawm, simplified versions. Koj tuaj yeem ua txhua tus neeg ua haujlwm hauv kev tswj hwm lossis kev txhawb nqa kev kawm, uas yog, koj cob qhia qhov ua si Atari games lossis qee yam zoo li ntawd lossis qee qhov kev ua si uas muaj qee qhov tsis paub tseeb hauv nws.
ZDNet: Ua tsaug rau koj lub sijhawm, Yann.
English
Arabic
Belarusian
Bengali
Bosnian
Bulgarian
Chinese (Simplified)
Croatian
Czech
Danish
Dutch
English
Estonian
Filipino
Finnish
French
Georgian
German
Greek
Gujarati
Hausa
Hebrew
Hindi
Hmong
Hungarian
Igbo
Indonesian
Italian
Japanese
Javanese
Kazakh
Korean
Kurdish (Kurmanji)
Kyrgyz
Lao
Latvian
Lithuanian
Macedonian
Malagasy
Norwegian
Persian
Polish
Portuguese
Punjabi
Romanian
Russian
Serbian
Slovak
Slovenian
Spanish
Swedish
Thai
Turkish
Ukrainian
Vietnamese
Yoruba