NeuroIPS () ke kopano e kholo ka ho fetisisa lefatšeng ea ho ithuta ka mochini le bohlale ba maiketsetso le ketsahalo e kholo lefats'eng la thuto e tebileng.
Na rona, baenjiniere ba DS, le rona re tla tseba biology, lipuo, le psychology lilemong tse leshome tse ncha? Re tla u bolella tlhahlobisong ea rona.
Selemong sena kopano e ile ea kopanya batho ba fetang 13500 ba tsoang linaheng tse 80 Vancouver, Canada. Ena hase selemo sa pele seo Sberbank e se emetseng Russia kopanong - sehlopha sa DS se buile ka ts'ebetsong ea ML lits'ebetsong tsa banka, ka tlhōlisano ea ML le ka bokhoni ba sethala sa Sberbank DS. Ke mekhoa efe ea mantlha ea 2019 sechabeng sa ML? Barupeluoa ba seboka ba re: и .
Selemong sena, NeurIPS e amohetse lipampiri tse fetang 1400-li-algorithms, mehlala e mecha, le lisebelisoa tse ncha ho data e ncha.
Tse ka Hare:
- Mekhoa
-
- Tlhaloso ea mohlala
- Multidisciplinarity
- Ho Bea Mabaka
- RL
- GAN
- Lipuo tsa Motheo tse Memang
-
- "Social Intelligence", Blaise Aguera y Arcas (Google)
- "Veridical Data Science", Bin Yu (Berkeley)
- "Ho Etsa Mohlala oa Boitšoaro ba Motho ka ho Ithuta ka Mochini: Menyetla le Liphephetso", Nuria M Oliver, Albert Ali Salah
- "Ho tloha ho System 1 ho ea ho Thuto e tebileng ea 2", Yoshua Bengio
Litloaelo tsa 2019
1. Tlhaloso ea mohlala le mokhoa o mocha oa ML
Sehlooho se seholo sa kopano ke tlhaloso le bopaki ba hore na ke hobane'ng ha re fumana liphello tse itseng. Motho a ka bua ka nako e telele ka bohlokoa ba filosofi ea tlhaloso ea "lebokose le letšo", empa ho ne ho e-na le mekhoa e mengata ea sebele le tsoelo-pele ea theknoloji sebakeng sena.
Mokhoa oa ho pheta-pheta mehlala le ho fumana tsebo ho tsoa ho eona ke sesebelisoa se secha sa mahlale. Mehlala e ka sebetsa e le sesebelisoa sa ho fumana tsebo e ncha le ho e leka, 'me mohato o mong le o mong oa ho lokisa, ho koetlisa le ho sebelisoa ha mohlala o tlameha ho etsoa hape.
Karolo e kholo ea likhatiso ha e nehetsoe kahong ea mehlala le lisebelisoa, empa mathata a ho netefatsa polokeho, pepeneneng le bonnete ba liphetho. Haholo-holo, ho hlahile molapo o fapaneng mabapi le litlhaselo tsa mohlala (litlhaselo tsa bahanyetsi), 'me likhetho tsa litlhaselo tse peli tsa koetliso le litlhaselo tsa kopo li nkoa.
Lingoliloeng:
- - sengoloa sa lenaneo mabapi le mokhoa oa ho netefatsa mohlala. E kenyelletsa kakaretso ea lisebelisoa tsa sejoale-joale tsa ho toloka mehlala, haholo ts'ebeliso ea tlhokomelo le ho fumana bohlokoa ba likarolo ka "distilling" neural network ka mefuta ea mela.
- 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
ExBert.net e bonts'a tlhaloso ea mohlala bakeng sa mesebetsi ea ho sebetsana le mongolo
2. Mekhoa e mengata
Ho netefatsa netefatso e tšepahalang le ho theha mekhoa ea ho netefatsa le ho holisa tsebo, re hloka litsebi tsa mafapha a amanang le tsona tse nang le bokhoni ka nako e le 'ngoe ho ML le sebakeng sa thuto (meriana, lipuo, neurobiology, thuto, joalo-joalo). Ho bohlokoa haholo ho hlokomela boteng ba bohlokoa ba mesebetsi le lipuo ho li-neuroscience le mahlale a kelello - ho na le kamano ea litsebi le ho alima mehopolo.
Ntle le kamano ena, ho ntse ho hlaha mekhoa e mengata ea ho sebetsana le tlhahisoleseling e tsoang mehloling e fapaneng: mongolo le linepe, mongolo le lipapali, li-database tsa graph + mongolo le linepe.
Lingoliloeng:
- Neuroscience + ML -
- VisualQA -
- RL + NLP -
Mefuta e 'meli - setsebi sa maano le molaoli - e ipapisitse le leano la ho bapala la inthanete la RL le NLP
3. Ho Bea Mabaka
Ho matlafatsa bohlale ba maiketsetso ke motsamao o lebisang lits'ebetsong tsa ho ithuta, "ho tseba", ho beha mabaka le ho beha mabaka. Haholo-holo, ho nts'etsapele maikutlo a causal le commonsense mabaka. Tse ling tsa litlaleho li reretsoe ho ithuta ka meta (mabapi le mokhoa oa ho ithuta ho ithuta) le motsoako oa theknoloji ea DL e nang le 1st le 2nd order logic - lentsoe Artificial General Intelligence (AGI) e fetoha lentsoe le tloaelehileng lipuong tsa libui.
Lingoliloeng:
- 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.Reinforcement Learning
Boholo ba mosebetsi o ntse o tsoela pele ho nts'etsapele libaka tsa setso tsa RL - DOTA2, Starcraft, ho kopanya meralo e nang le pono ea khomphutha, NLP, li-database tsa graph.
Letsatsi le arohaneng la kopano le ne le nehetsoe ho workshop ea RL, eo ho eona ho ileng ha hlahisoa meralo ea Optimistic Actor Critic Model, e phahametseng tsohle tse fetileng, haholo-holo Soft Actor Critic.
Lingoliloeng:
- ; Kamil Ciosek, Quan Vuong, Robert Loftin, Katja Hofmann
- ; Yasuhiro Fujita (Preferred Networks, Inc.)*; Toshiki Kataoka (Preferred Networks, Inc.); Prabhat Nagarajan (Likhokahano Tse Ratoang); Takahiro Ishikawa (University of Tokyo) [external pdf link].
- ; Danijar Hafner (Google)*; Timothy Lillicrap (DeepMind); Jimmy Ba (Univesithi ea Toronto); Mohammad Norouzi (Google Brain)
Libapali tsa StarCraft li loana le mohlala oa Alphastar (DeepMind)
5.GAN
Marang-rang a hlahisang a ntse a le leseling: mesebetsi e mengata e sebelisa vanilla GANs bakeng sa bopaki ba lipalo, hape e e sebelisa ka litsela tse ncha, tse sa tloaelehang (mefuta e hlahisang litšoantšo, ho sebetsa le letoto, kopo ea ho baka-le-phello likamano ho data, joalo-joalo).
Lingoliloeng:
- Sangwoo Mo, Chiheon Kim, Sungwoong Kim, Minsu Cho, Jinwoo Shin
- Dan Zhang, Anna Khoreva
- Lei Xu, Maria Skoularidou, Alfredo Cuesta-Infante, Kalyan Veeramachaneni
Kaha mosebetsi o mongata o ile oa amoheloa Ka tlase re tla bua ka lipuo tsa bohlokoa ka ho fetisisa.
Lipuo Tse Memetsoeng
"Social Intelligence", Blaise Aguera y Arcas (Google)
Puo e shebana le mokhoa o akaretsang oa ho ithuta ka mochini le litebello tsa ho fetola indasteri hona joale - re tobana le mathata afe? Boko le thuto ea ho iphetola ha lintho li sebetsa joang, hona ke hobane’ng ha re sa sebelise hakaalo seo re seng re ntse re se tseba ka tsoelo-pele ea litsamaiso tsa tlhaho?
Nts'etsopele ea indasteri ea ML haholo e tsamaellana le liketsahalo tsa bohlokoa tsa nts'etsopele ea Google, e phatlalatsang lipatlisiso tsa eona ho NeurIPS selemo le selemo:
- 1997 - ho qalisoa ha lisebelisoa tsa ho batla, li-server tsa pele, matla a manyenyane a k'homphieutha
- 2010 - Jeff Dean o qala morero oa Google Brain, boom ea marang-rang a neural qalong.
- 2015 - ts'ebetsong ea indasteri ea marang-rang a neural, ho lemoha sefahleho ka potlako ka ho toba mochine oa lehae, li-processor tsa boemo bo tlaase tse etselitsoeng lik'homphieutha tsa tensor - TPU. Google e qala Coral ai - analogue ea raspberry pi, komporo e nyane bakeng sa ho hlahisa marang-rang a neural lits'ebetsong tsa liteko.
- 2017 - Google e qala ho nts'etsapele koetliso e arohaneng le ho kopanya liphetho tsa koetliso ea marang-rang ea neural ho tsoa ho lisebelisoa tse fapaneng ho ea ho mohlala o le mong - ho Android.
Kajeno, indasteri eohle e inehetse ho ts'ireletso ea data, ho kopanya, le ho pheta-pheta liphetho tsa thuto ho lisebelisoa tsa lehae.
- tataiso ea ML moo mehlala ka bomong e ithutang e ikemetseng 'me e kopanngoa hore e be mohlala o le mong (ntle le ho beha data ea mohloli bohareng), e lokiselitsoeng liketsahalo tse sa tloaelehang, liphapang, botho, joalo-joalo. Lisebelisoa tsohle tsa Android ha e le hantle ke komporo e kholo ea komporo ea Google.
Mehlala ea tlhahiso e ipapisitseng le thuto e kopaneng ke tataiso e tšepisang nakong e tlang ho latela Google, e leng "mehatong ea pele ea kholo e kholo." Li-GAN, ho ea ka morupeli, li khona ho ithuta ho hlahisa boitšoaro ba bongata ba lintho tse phelang le mekhoa ea ho nahana.
U sebelisa mohlala oa meralo e 'meli e bonolo ea GAN, ho bonts'oa hore ho batla tsela ea optimization ho eona ho lelera ka selikalikoe, ho bolelang hore ts'ebetso e joalo ha e etsahale. Ka nako e ts'oanang, mehlala ena e atlehile haholo ho etsisa liteko tseo litsebi tsa baeloji li li etsang ka bongata ba libaktheria, li ba qobella ho ithuta mekhoa e mecha ea boitšoaro ha ba batla lijo. Re ka fihlela qeto ea hore bophelo bo sebetsa ka tsela e fapaneng le ts'ebetso ea optimization.
Ho tsamaea ka GAN Optimization
Sohle seo re se etsang moralong oa ho ithuta ka mochini hona joale ke mesebetsi e moqotetsane ebile e hlophisehile haholo, athe mekhoa ena ha e phethehe hantle ebile ha e tsamaellane le tsebo ea rona ea thuto libakeng tse kang neurophysiology le biology.
Seo e hlileng e leng sa bohlokoa ho alima lefapheng la neurophysiology haufinyane ke meralo e mecha ea neuron le tlhahlobo e nyane ea mekhoa ea ho khutlisa liphoso.
Boko ba motho ka bobona ha bo ithute joalo ka marang-rang a neural:
- Ha a na lintho tse ka sehloohong tse kenyang letsoho, ho kenyeletsoa le tse entsoeng ka kutlo le bongoaneng
- O na le litaelo tsa tlhaho tsa khōlo ea tlhaho (takatso ea ho ithuta puo ho tloha ho lesea, ho tsamaea hantle)
Ho koetlisa boko ba motho ka mong ke mosebetsi oa boemo bo tlaase; mohlomong re lokela ho nahana ka "likolone" tsa batho ba fetohang ka potlako ba fetisetsang tsebo ho ba bang ho hlahisa mekhoa ea ho iphetola ha lihlopha.
Seo re ka se nkang ho li-algorithms tsa ML hona joale:
- Sebelisa mefuta ea lisele tse netefatsang ho ithuta ha baahi, empa bophelo bo bokhutšoane ba motho ka mong ("boko ba motho ka mong")
- Ho ithuta ho fokolang ho sebelisa mehlala e fokolang
- Mehaho e rarahaneng ea li-neuron, mesebetsi ea ho kenya tšebetsong e fapaneng hanyane
- Ho fetisetsa "genome" melokong e tlang - backpropagation algorithm
- Hang ha re hokela neurophysiology le neural networks, re tla ithuta ho aha boko bo sebetsang ka bongata ho tsoa likarolong tse ngata.
Ho latela pono ena, ts'ebetso ea litharollo tsa SOTA e kotsi 'me e lokela ho nchafatsoa molemong oa ho etsa mesebetsi e tloaelehileng (benchmarks).
"Veridical Data Science", Bin Yu (Berkeley)
Tlaleho e shebane le bothata ba ho toloka mekhoa ea ho ithuta ka mochini le mokhoa oa ho etsa liteko le netefatso ea bona ka kotloloho. Mohlala ofe kapa ofe oa ML o koetlisitsoeng o ka nkoa e le mohloli oa tsebo e lokelang ho ntšoa ho oona.
Libakeng tse ngata, haholo-holo meriana, tšebeliso ea mohlala ha e khonehe ntle le ho ntša tsebo ena e patiloeng le ho hlalosa liphello tsa mohlala - ho seng joalo re ke ke ra ba le bonnete ba hore liphello li tla ba tse tsitsitseng, tse sa tloaelehang, tse ka tšeptjoang, 'me li ke ke tsa bolaea mokuli. Tataiso e felletseng ea mokhoa oa ho sebetsa e ntse e tsoela pele ka har'a paradigm e tebileng ea ho ithuta mme e fetela ka nģ'ane ho meeli ea eona - saense ea data ea veridical. Ke eng?
Re batla ho fihlela boleng bo joalo ba likhatiso tsa mahlale le ho ikatisa ha mefuta eo e leng:
- lebelloang
- computable
- tsitsitseng
Melao-motheo ena e meraro e theha motheo oa mokhoa o mocha. Mefuta ea ML e ka hlahlojoa joang khahlano le mekhoa ee? Tsela e bonolo ka ho fetisisa ke ho haha mefuta e fetolehang hang-hang (ho fokotseha, lifate tsa liqeto). Leha ho le joalo, re boetse re batla ho fumana melemo ea hang-hang ea ho ithuta ho tebileng.
Ho na le litsela tse 'maloa tsa ho sebetsana le bothata:
- toloka mohlala;
- sebelisa mekhoa e thehiloeng tlhokomelong;
- sebelisa li-ensembles tsa li-algorithms ha u koetlisa, 'me u netefatse hore mefuta e ka tolokang ka mela e ithuta ho bolela esale pele likarabo tse tšoanang le tsa marang-rang a marang-rang, likarolo tsa ho toloka ho tsoa moetsong oa mola;
- fetola le ho eketsa lintlha tsa koetliso. Sena se kenyelletsa ho eketsa lerata, tšitiso, le ho eketsa lintlha;
- mekhoa leha e le efe e thusang ho etsa bonnete ba hore liphetho tsa mohlala ha lia iketsahalla feela 'me ha li itšetlehe ka tšitiso e nyenyane e sa batleheng (litlhaselo tsa bahanyetsi);
- hlalosa mohlala ka mor'a 'nete, ka mor'a koetliso;
- ho ithuta ho hlahisa litekanyo ka litsela tse sa tšoaneng;
- ithuta menyetla ea likhopolo-taba tsohle, kabo ea lihlopha.
Tlhaselo ea sera
Liphoso tsa mohlala li bitsa motho e mong le e mong: mohlala o ka sehloohong ke mosebetsi oa Reinhart le Rogov."E ile ea susumetsa maano a moruo a linaha tse ngata tsa Europe mme ea ba qobella ho latela melaoana ea maemo a tlase, empa tlhahlobo e hlokolosi ea data le lilemo tsa ts'ebetso ea bona hamorao e bonts'itse sephetho se fapaneng!
Theknoloji efe kapa efe ea ML e na le potoloho ea eona ea bophelo ho tloha ts'ebetsong ho isa ts'ebetsong. Sepheo sa mokhoa o mocha ke ho hlahloba melao-motheo e meraro mohatong ka mong oa bophelo ba mohlala.
Kakaretso:
- Ho ntse ho etsoa merero e mengata e tla thusa mohlala oa ML hore o tšepahale haholoanyane. Sena ke, mohlala, deeptune (link to: );
- Bakeng sa nts'etsopele e eketsehileng ea mokhoa, hoa hlokahala ho ntlafatsa haholo boleng ba lingoliloeng tšimong ea ML;
- Ho ithuta ka mochini ho hloka baetapele ba nang le lithupelo tse fapaneng le boiphihlelo mafapheng a tekheniki le a botho.
"Ho Etsa Mohlala oa Boitšoaro ba Motho ka ho Ithuta ka Mochini: Menyetla le Liphephetso" Nuria M Oliver, Albert Ali Salah
Puo e inehetseng ho mohlala oa boitšoaro ba batho, metheo ea eona ea theknoloji le menyetla ea ts'ebeliso.
Mohlala oa boitšoaro ba motho o ka aroloa ka:
- boitšoaro ba motho ka mong
- boitšoaro ba sehlopha se senyane sa batho
- boitsoaro ba bongata
E 'ngoe le e' ngoe ea mefuta ena e ka etsoa mohlala ho sebelisoa ML, empa ka boitsebiso bo fapaneng ka ho feletseng le likarolo. Mofuta o mong le o mong o boetse o na le litaba tsa oona tsa boitšoaro tseo morero ka mong o fetang ho tsona:
- boitšoaro ba motho ka mong - bosholu ba boitsebiso, bo tebileng ba bohata;
- boitšoaro ba lihlopha tsa batho - de-anonymization, ho fumana tlhahisoleseding mabapi le metsamao, mehala ea thelefono, joalo-joalo;
boitšoaro ba motho ka mong
Haholo-holo e amanang le sehlooho sa Computer Vision - ho lemoha maikutlo a batho le liketso. Mohlomong feela ho latela moelelo oa taba, ka nako, kapa ka tekanyo e lekanyelitsoeng ea ho feto-fetoha ha maikutlo a hae. slide e bontša ho elelloa maikutlo a Mona Lisa a sebelisa moelelo oa maikutlo a basali ba Mediterranean. Sephetho: pososelo ea thabo, empa ka ho nyelisa le ho nyonya. Lebaka le ka etsahala haholo ka mokhoa oa botekgeniki oa ho hlalosa maikutlo a "sa nke lehlakore".
Boitšoaro ba sehlopha se senyane sa batho
Ho fihlela joale mohlala o mobe ka ho fetisisa o bakoa ke boitsebiso bo sa lekaneng. Ka mohlala, ho ile ha bontšoa mesebetsi ea 2018 - 2019. ho batho ba bangata X ba bangata ba livideo (bapisa. 100k++ datasets ea litšoantšo). Ho etsa mohlala o motle oa mosebetsi ona, ho hlokahala tlhaiso-leseling ea multimodal, haholo-holo ho tsoa ho li-sensor tsa altimeter ea 'mele, thermometer, ho rekota maekrofono, jj.
Boitšoaro ba bongata
Sebaka se tsoetseng pele ka ho fetisisa, kaha moreki ke UN le linaha tse ngata. Lik'hamera tsa ho shebella ka ntle, lintlha tse tsoang litora tsa thelefono - litefiso, li-SMS, li-call, data mabapi le motsamao pakeng tsa meeli ea naha - sena sohle se fana ka setšoantšo se ka tšeptjoang haholo sa motsamao oa batho le ho hloka botsitso sechabeng. Lisebelisoa tse ka khonehang tsa thekenoloji: ho ntlafatsa ts'ebetso ea pholoso, thuso le ho tlosoa ha baahi ka nako nakong ea maemo a tšohanyetso. Mefuta e sebelisitsoeng hangata e ntse e hlalosoa hampe - tsena ke li-LSTM tse fapaneng le marang-rang a convolutional. Ho bile le polelo e khuts'oane ea hore UN e ne e batla molao o mocha o tla tlama likhoebo tsa Europe ho arolelana lintlha tse sa tsejoeng tse hlokahalang bakeng sa lipatlisiso life kapa life.
"Ho tloha ho System 1 ho ea ho Thuto e tebileng ea 2", Yoshua Bengio
Thutong ea Joshua Bengio, thuto e tebileng e kopana le neuroscience boemong ba ho ipehela lipakane.
Bengio e khetholla mefuta e 'meli ea mathata ho latela mokhoa oa mohapi oa Khau ea Nobel Daniel Kahneman (buka "")
mofuta oa 1 - Tsamaiso ea 1, liketso tse sa tsebeng letho tseo re li etsang "ka tsela e iketsang" (boko ba boholo-holo): ho khanna koloi libakeng tse tloaelehileng, ho tsamaea, ho lemoha lifahleho.
mofuta oa 2 - Sisteme ea 2, liketso tse hlokolosi (cerebral cortex), ho beha sepheo, tlhahlobo, monahano, mesebetsi e kopaneng.
AI ho fihlela joale e se e fihlile bophahamong bo lekaneng feela mesebetsing ea mofuta oa pele, ha mosebetsi oa rona e le ho e tlisa ho ea bobeli, ho e ruta ho etsa mesebetsi e mengata le ho sebetsa ka mokhoa o utloahalang le tsebo e phahameng ea kutloisiso.
Ho fihlela sepheo sena ho sisintswe:
- mesebetsing ea NLP, sebelisa tlhokomelo e le mokhoa oa bohlokoa oa ho nahana ka mohlala
- sebelisa thuto ea meta-thuto le boemeli ho ntlafatsa likarolo tsa mohlala tse susumetsang kelello le tikoloho ea tsona - 'me motheong oa tsona li tsoela pele ho sebetsa ka mehopolo ea maemo a holimo.
Sebakeng sa sephetho, mona ke puo e memetsoeng: Bengio ke e mong oa bo-ramahlale ba bangata ba lekang ho holisa tšimo ea ML ho feta mathata a optimization, SOTA le meralo e mecha.
Potso e ntse e bulehile hore na ho kopana ha mathata a kelello, tšusumetso ea puo ea ho nahana, neurobiology le algorithms ke eng e re letetseng nakong e tlang mme e tla re lumella ho fallela mechine e "nahanang" joaloka batho.
Спасибо!
Source: www.habr.com