I-NeuroIPS () ingqungquthela enkulu kunazo zonke emhlabeni yokufunda ngomshini nobuhlakani bokwenziwa kanye nomcimbi oyinhloko emhlabeni wokufunda okujulile.
Ngabe thina, onjiniyela be-DS, sizophinde sibe yingcweti yebhayoloji, izilimi, kanye nesayikholoji kuleli shumi leminyaka elisha? Sizokutshela ekubuyekezeni kwethu.
Kulo nyaka le ngqungquthela ihlanganise abantu abangaphezu kuka-13500 80 abavela emazweni angu-2019 eVancouver, eCanada. Lona akuwona unyaka wokuqala ukuthi i-Sberbank imele iRussia engqungqutheleni - ithimba le-DS likhulume ngokuqaliswa kwe-ML ezinkambisweni zebhange, mayelana nokuncintisana kwe-ML kanye namakhono esiteji se-Sberbank DS. Iziphi izinkambiso eziphambili zango-XNUMX emphakathini we-ML? Abahlanganyeli benkomfa bathi: и .
Kulo nyaka, i-NeurIPS yamukele amaphepha angaphezu kwe-1400—ama-algorithms, amamodeli amasha, nezinhlelo zokusebenza ezintsha kudatha entsha.
Okuqukethwe:
- Amathrendi
-
- Imodeli yokutolika
- Ukuziphatha Okuningi
- Ukubonisana
- RL
- GAN
- Izinkulumo Ezimenyiwe Eziyisisekelo
-
- "Social Intelligence", uBlaise Aguera y Arcas (Google)
- "Veridical Data Science", uBin Yu (Berkeley)
- "Ukumodela Ukuziphatha Komuntu Ngokufunda Ngomshini: Amathuba Nezinselele", uNuria M Oliver, u-Albert Ali Salah
- "Kusuka kuSistimu 1 ukuya kuSistimu 2 Ukufunda Okujulile", u-Yoshua Bengio
Amathrendi we-2019 Wonyaka
1. Imodeli yokutolika kanye nendlela entsha ye-ML
Isihloko esikhulu sengqungquthela ukutolika kanye nobufakazi bokuthi kungani sithola imiphumela ethile. Umuntu angakhuluma isikhathi eside ngokubaluleka kwefilosofi yokuhumusha "ibhokisi elimnyama", kodwa kwakukhona izindlela zangempela kanye nentuthuko yezobuchwepheshe kule ndawo.
Indlela yokwenza yokuphindaphinda amamodeli nokukhipha ulwazi kuwo iyikhithi yamathuluzi entsha yesayensi. Amamodeli angasebenza njengethuluzi lokuthola ulwazi olusha kanye nokuluvivinya, futhi isigaba ngasinye sokucutshungulwa kwangaphambili, ukuqeqeshwa kanye nokusetshenziswa kwemodeli kufanele kuphindaphindeke.
Ingxenye ebalulekile yokushicilelwa ayinikezelwe ekwakhiweni kwamamodeli namathuluzi, kodwa ezinkingeni zokuqinisekisa ukuphepha, ukucaca nokuqinisekiswa kwemiphumela. Ikakhulukazi, kuvele umfudlana ohlukile mayelana nokuhlaselwa kwemodeli (ukuhlaselwa kwezitha), futhi izinketho zakho kokubili ukuhlaselwa kokuqeqeshwa nokuhlaselwa kwesicelo kucatshangelwa.
Izindatshana:
- - i-athikili yohlelo mayelana nendlela yokuqinisekisa imodeli. Kufaka phakathi ukubuka konke kwamathuluzi esimanje amamodeli okuhumusha, ikakhulukazi, ukusetshenziswa kokunaka kanye nokuthola ukubaluleka kwesici “ngokukhipha” inethiwekhi ye-neural ngamamodeli alayini.
- Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, Jonathan K. Su
- Sara Hooker, Dumitru Erhan, Pieter-Jan Kindermans, Been Kim
- Alexander Mott, Daniel Zoran, Mike Chrzanowski, Daan Wierstra, Danilo Jimenez Rezende
- Xiao Li, Yu Wang, Sumanta Basu, Karl Kumbier, Bin Yu
- Jaemin Yoo, Minyong Cho, Taebum Kim, U Kang
- Edward Raff
I-ExBert.net ibonisa ukutolika okuyimodeli yemisebenzi yokucubungula umbhalo
2. Ukwenza izinto eziningi
Ukuqinisekisa ukuqinisekiswa okuthembekile nokuthuthukisa izindlela zokuqinisekisa nokwandisa ulwazi, sidinga ochwepheshe emikhakheni ehlobene abanekhono kanye kanye ku-ML kanye nendawo yesifundo (umuthi, izilimi, i-neurobiology, imfundo, njll.). Kubaluleke kakhulu ukuqaphela ukuba khona okubaluleke kakhulu kwemisebenzi nezinkulumo kuma-neuroscience kanye nesayensi yengqondo - kukhona ukuhlangana kochwepheshe kanye nokubolekwa kwemibono.
Ngaphezu kwalokhu kusondelana, kuvela izinhlobonhlobo eziningi ekucutshungulweni okuhlanganyelwe kolwazi oluvela emithonjeni ehlukahlukene: umbhalo nezithombe, umbhalo nemidlalo, isizindalwazi segrafu + umbhalo nezithombe.
Izindatshana:
- I-Neuroscience + ML —
- I-VisualQA -
- I-RL + NLP -
Amamodeli amabili - amasu kanye nesiphathimandla - asuselwa ku-RL ne-NLP yokudlala isu eliku-inthanethi
3. Ukubonisana
Ukuqinisa ubuhlakani bokwenziwa wumnyakazo obheke ezinhlelweni zokuzifundela, "ukuqaphela", ukucabanga nokucabanga. Ikakhulukazi, i-causal inference kanye ne-commonsense yokucabanga iyathuthuka. Eminye yemibiko igxile ekufundeni imeta (mayelana nendlela yokufunda ukufunda) kanye nenhlanganisela yobuchwepheshe be-DL enomqondo woku-1 nowesibili we-oda - igama elithi Artificial General Intelligence (AGI) seliba yitemu elivamile ezinkulumweni zezikhulumi.
Izindatshana:
- Weijiang Yu, Jingwen Zhou, Weihao Yu, Xiaodan Liang, Nong Xiao
- Wang-Zhou Dai, Qiuling Xu, Yang Yu, Zhi-Hua Zhou
- Vaishak Belle, Brendan Juba
- Anton Bakhtin, Laurens van der Maaten, Justin Johnson, Laura Gustafson, Ross Girshick
- Dinesh Garg, Shajith Ikbal, Santosh K. Srivastava, Harit Vishwakarma, Hima Karanam, L Venkata Subramaniam
4.Ukugcizelela Ukufunda
Iningi lomsebenzi liyaqhubeka nokuthuthukisa izindawo zendabuko ze-RL - DOTA2, i-Starcraft, ukuhlanganisa izakhiwo ezinombono wekhompyutha, i-NLP, i-graph database.
Usuku oluhlukile lwengqungquthela lunikezelwe ku-workshop ye-RL, lapho kwethulwa khona ukwakhiwa kwe-Optimistic Actor Critic Model, okudlula zonke ezedlule, ikakhulukazi i-Soft Actor Critic.
Izindatshana:
- ; Kamil Ciosek, Quan Vuong, Robert Loftin, Katja Hofmann
- ; Yasuhiro Fujita (Preferred Networks, Inc.)*; Toshiki Kataoka (Preferred Networks, Inc.); Prabhat Nagarajan (Amanethiwekhi Akhethwayo); Takahiro Ishikawa (The University of Tokyo) [external pdf link].
- ; U-Danijar Hafner (Google)*; Timothy Lillicrap (DeepMind); Jimmy Ba (University of Toronto); U-Mohammad Norouzi (Google Brain)
Abadlali be-StarCraft balwa nemodeli ye-Alphastar (DeepMind)
5.GAN
Amanethiwekhi akhiqizayo asabonakala: imisebenzi eminingi isebenzisa ama-vanilla GAN ukuze uthole ubufakazi bezibalo, futhi iwasebenzisa ngezindlela ezintsha, ezingajwayelekile (amamodeli akhiqiza igrafu, ukusebenza ngochungechunge, isicelo sokudala ubudlelwano nomthelela kudatha, njll.).
Izindatshana:
- Sangwoo Mo, Chiheon Kim, Sungwoong Kim, Minsu Cho, Jinwoo Shin
- Dan Zhang, Anna Khoreva
- Lei Xu, Maria Skoularidou, Alfredo Cuesta-Infante, Kalyan Veeramachaneni
Njengoba umsebenzi owengeziwe wamukelwa Ngezansi sizokhuluma ngezinkulumo ezibaluleke kakhulu.
Izinkulumo Ezimenyiwe
"Social Intelligence", uBlaise Aguera y Arcas (Google)
Inkulumo igxile endleleni ejwayelekile yokufunda komshini kanye namathemba okushintsha imboni njengamanje - yiziphi izimpambano-mgwaqo esibhekene nazo? Kusebenza kanjani ubuchopho nokuziphendukela kwemvelo, futhi kungani singakusebenzisi kangako lokho esesikwazi kakade mayelana nokuthuthukiswa kwezimiso zemvelo?
Ukuthuthukiswa kwezimboni kwe-ML kuhambisana kakhulu nezigigaba zokuthuthuka kwe-Google, eshicilela ucwaningo lwayo nge-NeurIPS unyaka nonyaka:
- 1997 - ukwethulwa kwezikhungo zokusesha, amaseva okuqala, amandla amancane wekhompyutha
- 2010 - UJeff Dean wethula iphrojekthi ye-Google Brain, ukuchuma kwamanethiwekhi e-neural ekuqaleni.
- 2015 - ukuqaliswa kwezimboni kwamanethiwekhi e-neural, ukubonwa kobuso okusheshayo ngokuqondile kudivayisi yasendaweni, amaphrosesa asezingeni eliphansi enzelwe i-tensor computing - TPU. I-Google yethula i-Coral ai - i-analogue ye-raspberry pi, ikhompyutha encane yokwethula amanethiwekhi e-neural ekufakweni kokuhlola
- 2017 - I-Google iqala ukuthuthukisa ukuqeqeshwa okuhlukaniswe futhi ihlanganise imiphumela yokuqeqeshwa kwenethiwekhi ye-neural kusuka kumadivayisi ahlukene ibe yimodeli eyodwa - ku-Android
Namuhla, yonke imboni izinikele ekuvikelekeni kwedatha, ukuhlanganisa, nokuphindaphinda imiphumela yokufunda kumadivayisi asendaweni.
- isiqondiso se-ML lapho amamodeli ngamanye afunda khona ngokuzimela abese ehlanganiswa abe yimodeli eyodwa (ngaphandle kokuhlanganisa idatha yomthombo), elungiselwe imicimbi engavamile, okudidayo, ukwenza kube ngokwakho, njll. Wonke amadivayisi e-Android empeleni ayikhompyutha eyi-computing eyodwa ye-Google.
Amamodeli akhiqizayo asuselwe ekufundeni okuhlanganyelwe ayisiqondiso sesikhathi esizayo esithembisayo ngokusho kwe-Google, "esezigabeni zokuqala zokukhula okukhulu." Ama-GAN, ngokusho komfundisi, ayakwazi ukufunda ukukhiqiza kabusha ukuziphatha kwenqwaba yezinto eziphilayo nama-algorithms okucabanga.
Kusetshenziswa isibonelo sezakhiwo ezimbili ezilula ze-GAN, kuboniswa ukuthi kuzo ukusesha kwendlela yokuthuthukisa kuzulazula embuthanweni, okusho ukuthi ukwenza kahle kanjalo akwenzeki. Ngesikhathi esifanayo, lawa mamodeli aphumelela kakhulu ekulingiseni ukuhlola okwenziwa yizazi zebhayoloji emiphakathini yamagciwane, okubaphoqa ukuba bafunde amasu amasha okuziphatha lapho befuna ukudla. Singaphetha ngokuthi ukuphila kusebenza ngendlela ehlukile kunomsebenzi wokulungiselela.
Ukuhamba Ukuthuthukisa i-GAN
Konke esikwenzayo ohlakeni lokufunda komshini manje kuyimisebenzi emincane futhi esemthethweni ngokwedlulele, kuyilapho lezi zindlela ezihlelekile azihlanganisi kahle futhi azihambisani nolwazi lwesihloko sethu ezindaweni ezifana ne-neurophysiology ne-biology.
Okufanelekile ngempela ukuboleka emkhakheni we-neurophysiology esikhathini esizayo esiseduze izakhiwo ezintsha ze-neuron kanye nokubuyekezwa okuncane kwezinqubo zokusabalalisa amaphutha emuva.
Ubuchopho bomuntu ngokwawo abufundi njengenethiwekhi ye-neural:
- Akanayo imibono eyisisekelo engahleliwe, kuhlanganise naleyo ehlelwe ngezinzwa nasebuntwaneni
- Unezikhombisi-ndlela zemvelo zokukhula komzwelo (isifiso sokufunda ulimi kusukela enganeni, ehamba eqondile)
Ukuqeqesha ubuchopho bomuntu ngamunye kuwumsebenzi osezingeni eliphansi; mhlawumbe kufanele sicabangele “amakholoni” abantu abashintsha ngokushesha abadlulisela ulwazi komunye nomunye ukukhiqiza kabusha izindlela zokuziphendukela kweqembu.
Lokho esingakusebenzisa kuma-algorithms e-ML manje:
- Sebenzisa amamodeli omugqa wamaseli aqinisekisa ukufunda kwenani labantu, kodwa impilo emfushane yomuntu ngamunye (“ubuchopho bomuntu ngamunye”)
- Ukufunda okumbalwa kusetshenziswa izibonelo ezimbalwa
- Izakhiwo ze-neuron eziyinkimbinkimbi, imisebenzi yokuvula ehluke kancane
- Ukudlulisela "i-genome" ezizukulwaneni ezilandelayo - i-algorithm ye-backpropagation
- Uma sesixhumanisa i-neurophysiology kanye namanethiwekhi e-neural, sizofunda ukwakha ubuchopho obusebenzayo ezingxenyeni eziningi.
Kusukela kulo mbono, umkhuba wezixazululo ze-SOTA uyingozi futhi kufanele ubuyekezwe ukuze kuthuthukiswe imisebenzi evamile (izilinganiso).
"Veridical Data Science", uBin Yu (Berkeley)
Umbiko ugxile enkingeni yokutolika amamodeli okufunda omshini kanye nendlela yokuhlola kwawo okuqondile nokuqinisekisa. Noma iyiphi imodeli ye-ML eqeqeshiwe ingabonwa njengomthombo wolwazi okudingeka lukhishwe kuwo.
Ezindaweni eziningi, ikakhulukazi kwezokwelapha, ukusetshenziswa kwemodeli akunakwenzeka ngaphandle kokukhipha lolu lwazi olufihliwe nokuhumusha imiphumela yemodeli - ngaphandle kwalokho ngeke siqiniseke ukuthi imiphumela izoba ezinzile, engahleliwe, ethembekile, futhi ngeke ibulale isiguli. Yonke inkombandlela yendlela yokusebenza iyathuthuka ngaphakathi kwepharadigm yokufunda ejulile futhi idlulela ngale kwemingcele yayo - isayensi yedatha eqondile. Yini?
Sifuna ukuzuza ikhwalithi enjalo yokushicilelwa kwesayensi nokukhiqizwa kabusha kwamamodeli ayi:
- ukubikezelwa
- kuyasebenziseka
- ezinzile
Le migomo emithathu yakha isisekelo sendlela yokusebenza entsha. Angahlolwa kanjani amamodeli e-ML ngokumelene nalezi zindlela zokunquma? Indlela elula ukwakha amamodeli ahunyushwa ngokushesha (ukuhlehla, izihlahla zokunquma). Nokho, sifuna futhi ukuthola izinzuzo ezisheshayo zokufunda ngokujulile.
Izindlela ezimbalwa ezikhona zokwenza le nkinga:
- chaza imodeli;
- sebenzisa izindlela ezisekelwe ekunakeni;
- sebenzisa ama-algorithms ahlanganisiwe lapho uqeqeshwa, futhi uqinisekise ukuthi amamodeli ahunyushwa ngomugqa afunda ukubikezela izimpendulo ezifanayo njengenethiwekhi ye-neural, izici zokuhumusha ezivela kumodeli yomugqa;
- shintsha futhi uthuthukise idatha yokuqeqeshwa. Lokhu kuhlanganisa ukungeza umsindo, ukuphazamiseka, nokwandisa idatha;
- noma yiziphi izindlela ezisiza ukuqinisekisa ukuthi imiphumela yemodeli ayiyona into engahleliwe futhi ayixhomeki ekuphazamisekeni okuncane okungadingeki (ukuhlaselwa kwezitha);
- chaza imodeli ngemuva kweqiniso, ngemuva kokuqeqeshwa;
- ukutadisha izici izisindo ngezindlela ezihlukahlukene;
- funda amathuba awo wonke ama-hypotheses, ukusatshalaliswa kwekilasi.
Ukuhlasela kwezitha
Amaphutha wokumodela abiza wonke umuntu: isibonelo esihle umsebenzi kaReinhart noRogov.""Kuthonye izinqubomgomo zezomnotho zamazwe amaningi aseYurophu futhi kwawaphoqa ukuthi aphishekele izinqubomgomo zokunciphisa izindleko, kodwa ukuhlola ngokucophelela idatha kanye neminyaka yokucubungula kwayo kamuva kwabonisa umphumela ohlukile!
Noma ibuphi ubuchwepheshe be-ML bunomjikelezo wabo wokuphila kusukela ekusetshenzisweni kuya ekusetshenzisweni. Umgomo wendlela yokusebenza entsha ukuhlola izimiso ezintathu eziyisisekelo esigabeni ngasinye sempilo yemodeli.
Imiphumela:
- Amaphrojekthi amaningana ayathuthukiswa azosiza imodeli ye-ML ukuthi ithembeke kakhulu. Lokhu, ngokwesibonelo, i-deeptune (isixhumanisi ku: );
- Ukuze uthole ukuthuthukiswa okuqhubekayo kwendlela yokusebenza, kuyadingeka ukuthuthukisa kakhulu ikhwalithi yokushicilelwa emkhakheni we-ML;
- Ukufunda ngomshini kudinga abaholi abanokuqeqeshwa kwemikhakha eminingi nobungcweti kuyo yomibili imikhakha yezobuchwepheshe neyesintu.
"Ukumodela Ukuziphatha Komuntu Ngokufunda Ngomshini: Amathuba Nezinselelo" Nuria M Oliver, Albert Ali Salah
Isifundo esinikezelwe ekumodeleni ukuziphatha komuntu, izisekelo zakhona zobuchwepheshe kanye namathemba okusebenza.
Imodeli yokuziphatha komuntu ingahlukaniswa:
- ukuziphatha komuntu ngamunye
- ukuziphatha kweqembu elincane labantu
- ukuziphatha kwabantu abaningi
Ngalunye lwalezi zinhlobo lungamodelwa kusetshenziswa i-ML, kodwa ngolwazi lokufakwayo oluhluke ngokuphelele nezici. Uhlobo ngalunye luphinde lube nezindaba zalo zokuziphatha iphrojekthi ngayinye ehamba kuzo:
- ukuziphatha komuntu ngamunye - ukweba identity, deepfake;
- ukuziphatha kwamaqembu abantu - de-anonymization, ukuthola ulwazi mayelana nokunyakaza, izingcingo, njll;
ukuziphatha komuntu ngamunye
Ikakhulukazi ihlobene nesihloko se-Computer Vision - ukuqashelwa kwemizwa yabantu nokusabela. Mhlawumbe kuphela ngokomongo, ngokuhamba kwesikhathi, noma ngezinga elihlobene lokuhlukahluka kwemizwa yakhe. Isilayidi sibonisa ukuqashelwa kwemizwa ka-Mona Lisa sisebenzisa umongo ovela kububanzi bemizwa yabesifazane baseMedithera. Umphumela: ukumamatheka kwenjabulo, kodwa ngokudelela nokunengeka. Isizathu singenzeka kakhulu endleleni yobuchwepheshe yokuchaza umzwelo "ongathathi hlangothi".
Ukuziphatha kweqembu elincane labantu
Kuze kube manje imodeli embi kakhulu ingenxa yolwazi olunganele. Njengesibonelo, kuboniswe imisebenzi evela ku-2018 - 2019. kubantu abaningi X inqwaba yamavidiyo (cf. 100k++ amasethi edatha ezithombe). Ukumodela kahle lo msebenzi, kudingeka ulwazi lwe-multimodal, okungcono kakhulu oluvela kuzinzwa ku-altimeter yomzimba, ithemometha, ukuqoshwa kwemakrofoni, njll.
Ukuziphatha kwabantu abaningi
Indawo ethuthuke kakhulu, njengoba ikhasimende liyi-UN kanye nezifunda eziningi. Amakhamera okuqapha angaphandle, idatha evela emibhoshongweni yocingo - ukukhokhiswa, i-SMS, izingcingo, idatha yokunyakaza phakathi kwemingcele yombuso - konke lokhu kunikeza isithombe esinokwethenjelwa kakhulu sokunyakaza kwabantu nokungazinzi komphakathi. Ukusetshenziswa okungaba khona kobuchwepheshe: ukwenziwa kahle kwemisebenzi yokuhlenga, usizo kanye nokukhishwa okufika ngesikhathi kwabantu ngesikhathi sezimo eziphuthumayo. Amamodeli asetshenzisiwe awakahunyushwa kabi - lawa ngama-LSTM ahlukahlukene kanye namanethiwekhi okuxhumana. Kube nokuphawula okufushane kokuthi i-UN ibifuna umthetho omusha ozophoqa amabhizinisi ase-Europe ukuthi abelane ngedatha engaziwa edingekayo kunoma yiluphi ucwaningo.
"Kusuka kuSistimu 1 ukuya kuSistimu 2 Ukufunda Okujulile", u-Yoshua Bengio
Enkulumweni ka-Joshua Bengio, ukufunda okujulile kuhlangana ne-neuroscience ezingeni lokubeka imigomo.
U-Bengio uhlonza izinhlobo ezimbili eziyinhloko zezinkinga ngokwendlela yokwenza yomklomelo kaNobel uDaniel Kahneman (incwadi “")
uhlobo 1 - Isistimu 1, izenzo eziqulekile esizenzayo "ngokuzenzakalelayo" (ubuchopho basendulo): ukushayela imoto ezindaweni ezijwayelekile, ukuhamba ngezinyawo, ukubona ubuso.
uhlobo 2 - Isistimu 2, izenzo eziqaphelayo (i-cerebral cortex), ukubeka umgomo, ukuhlaziya, ukucabanga, imisebenzi eyinhlanganisela.
I-AI kuze kube manje isifinyelele ukuphakama okwanele kuphela emisebenzini yohlobo lokuqala, kanti umsebenzi wethu uwukuyisa kowesibili, ukuyifundisa ukwenza imisebenzi ehlukahlukene futhi isebenze ngokunengqondo kanye namakhono aphezulu okuqonda.
Ukuze kufinyelelwe lo mgomo kuhlongozwa:
- emisebenzini ye-NLP, sebenzisa ukunaka njengendlela eyinhloko yokumodela ukucabanga
- sebenzisa i-meta-learning nokufunda ngokumelela ukwenza imodeli engcono yezici ezithonya ukwazi nokwenza kwasendaweni - futhi ngokwesisekelo sazo uqhubekele ekusebenzeni ngemiqondo yezinga eliphezulu.
Esikhundleni sesiphetho, nansi inkulumo emenyiwe: U-Bengio ungomunye wososayensi abaningi abazama ukwandisa umkhakha we-ML ngale kwezinkinga zokuthuthukisa, i-SOTA kanye nezakhiwo ezintsha.
Umbuzo usalokhu uvulekile ukuthi inhlanganisela yezinkinga zokuqaphela, ithonya lolimi ekucabangeni, i-neurobiology kanye ne-algorithms ingakanani esilindele esikhathini esizayo futhi izosivumela ukuthi sithuthele emishinini "ecabanga" njengabantu.
Siyabonga!
Source: www.habr.com