NeuroIPS () shine babban taro na duniya akan koyo na'ura da basirar wucin gadi kuma babban taron a duniyar ilmantarwa mai zurfi.
Shin mu, injiniyoyin DS, mu ma, za mu iya ƙware kan ilimin halitta, ilimin harshe, da ilimin halin ɗan adam a cikin sabbin shekaru goma? Za mu gaya muku a cikin sharhinmu.
A bana taron ya hada fiye da mutane 13500 daga kasashe 80 a birnin Vancouver na kasar Canada. Wannan ba shine shekara ta farko da Sberbank ya wakilci Rasha a taron ba - ƙungiyar DS ta yi magana game da aiwatar da ML a cikin tsarin banki, game da gasar ML da kuma damar da Sberbank DS ke da shi. Menene manyan abubuwan da suka faru na 2019 a cikin al'ummar ML? Mahalarta taron sun ce: и .
A wannan shekara, NeurIPS ta karɓi fiye da takardu 1400-algorithms, sabbin samfura, da sabbin aikace-aikace zuwa sabbin bayanai.
Abubuwan:
- Juyawa
-
- Fassarar samfurin
- Multidisciplinarity
- Tunani
- RL
- GAN
- Babban Gayyatar Tattaunawa
-
- "Intelligence Social", Blaise Aguera da Arcas (Google)
- "Kimiyyar Bayanai ta Gaskiya", Bin Yu (Berkeley)
- "Tsarin Halin Dan Adam tare da Koyan Injin: Dama da Kalubale", Nuria M Oliver, Albert Ali Salah
- "Daga Tsarin 1 zuwa Tsarin 2 Zurfafa Koyo", Yoshua Bengio
Trends 2019
1. Fassarar samfuri da sabuwar hanyar ML
Babban batu na taron shine fassarar da shaida dalilin da yasa muke samun wasu sakamako. Mutum na iya yin magana na dogon lokaci game da mahimmancin falsafar fassarar "akwatin baki", amma akwai ƙarin hanyoyin gaske da ci gaban fasaha a wannan yanki.
Hanyar yin kwafin samfuri da kuma fitar da ilimi daga gare su sabon kayan aiki ne na kimiyya. Samfura na iya zama kayan aiki don samun sabon ilimi da gwada shi, kuma kowane mataki na gaba-gaba, horo da aikace-aikacen samfurin dole ne a sake yin su.
Yawancin wallafe-wallafen ba a sadaukar da su ga gina samfura da kayan aiki ba, amma ga matsalolin tabbatar da tsaro, bayyana gaskiya da tabbatar da sakamako. Musamman ma, wani rafi daban ya bayyana game da hare-hare a kan samfurin (hare-haren abokan gaba), kuma ana la'akari da zaɓuɓɓukan duka hare-hare akan horo da harin kan aikace-aikacen.
Labarai:
- - labarin shirye-shirye akan hanyoyin tabbatar da samfur. Ya haɗa da bayyani na kayan aikin zamani don fassarar ƙira, musamman, yin amfani da hankali da samun mahimmancin fasali ta hanyar “distilling” cibiyar sadarwar jijiya tare da ƙirar layi.
- 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 yana nuna fassarar samfuri don ayyukan sarrafa rubutu
2. Multidisciplinarity
Don tabbatar da ingantaccen tabbaci da haɓaka hanyoyin tabbatarwa da faɗaɗa ilimi, muna buƙatar ƙwararrun ƙwararru a fannoni masu alaƙa waɗanda lokaci guda suna da ƙwarewa a cikin ML da kuma a cikin yanki (magani, ilimin harshe, neurobiology, ilimi, da sauransu). Yana da mahimmanci a lura da kasancewar ayyukan da jawabai a cikin neurosciences da ilimin kimiyya - akwai kusantar ƙwararrun ƙwararru da karɓar ra'ayoyi.
Bugu da ƙari ga wannan kusanci, multidisciplinarity yana fitowa a cikin haɗin gwiwar sarrafa bayanai daga sassa daban-daban: rubutu da hotuna, rubutu da wasanni, bayanan jadawali + rubutu da hotuna.
Labarai:
- Kimiyyar Neuroscience + ML -
- VisualQA -
- RL + NLP -
Samfura guda biyu - strategist da zartarwa - dangane da dabarun wasan RL da NLP akan layi
3. Yin Tunani
Ƙarfafa hankali na wucin gadi motsi ne zuwa tsarin ilmantarwa, "sani", tunani da tunani. Musamman ma, dalilan da suka haifar da dalilai na fahimtar juna suna tasowa. Wasu daga cikin rahotannin sun keɓe ga ilimin meta (game da yadda ake koyan koyo) da haɗin fasahar DL tare da dabaru na 1st da 2nd - kalmar Artificial General Intelligence (AGI) ta zama kalma gama gari a cikin jawaban masu magana.
Labarai:
- 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.Karfafa Koyo
Yawancin ayyukan suna ci gaba da haɓaka wuraren gargajiya na RL - DOTA2, Starcraft, haɗa gine-gine tare da hangen nesa na kwamfuta, NLP, bayanan hoto.
An keɓe wata rana ta daban na taron don taron bita na RL, inda aka gabatar da tsarin gine-ginen ƙwararrun ƙwararrun ƙwararrun ƙwararrun ƙwararrun ƙwararrun ƙwararrun ƙwararrun ƙwararrun ƙwararrun ƙwararrun ƙwararrun ƙwararrun ƙwararrun ƙwararrun ƙwararrun ƙwararrun ƙwararrun ƙwararrun ƙwararrun ƙwararrun ƙwararrun ƙwararrun ƙwararrun ƙwararrun ƙwararrun ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararru.
Labarai:
- ; Kamil Ciosek, Quan Vuong, Robert Loftin, Katja Hofmann
- ; Yasuhiro Fujita (Networks Preferred, Inc.)*; Toshiki Kataoka (Networks da aka Fi so, Inc.); Prabhat Nagarajan (Cibiyoyin Sadarwar da Aka Fi so); Takahiro Ishikawa (Jami'ar Tokyo) [haɗin pdf na waje].
- ; Danjar Hafner (Google)*; Timothy Lillicrap (DeepMind); Jimmy Ba (Jami'ar Toronto); Mohammad Norouzi (Google Brain)
'Yan wasan StarCraft suna yaƙi da ƙirar Alphastar (DeepMind)
5.GAN
Cibiyoyin hanyoyin sadarwa na zamani har yanzu suna cikin tabo: ayyuka da yawa suna amfani da vanilla GANs don hujjojin lissafi, kuma suna amfani da su a cikin sabbin, hanyoyin da ba a saba gani ba (samfurin ƙira, aiki tare da jerin, aikace-aikacen haifar-da-sakamako dangantaka a cikin bayanai, da sauransu).
Labarai:
- Sangwoo Mo, Chiheon Kim, Sungwoong Kim, Minsu Cho, Jinwoo Shin
- Dan Zhang, Anna Khorva
- Lei Xu, Maria Skoularidou, Alfredo Cuesta-Infante, Kalyan Veeramachaneni
Tunda an sami ƙarin aiki A ƙasa za mu yi magana game da jawabai mafi mahimmanci.
Tattaunawar da Aka Gayyata
"Intelligence Social", Blaise Aguera da Arcas (Google)
Jawabin ya mayar da hankali ne kan tsarin koyan na'ura gabaɗaya da kuma hasashen da ake da shi na sauya masana'antu a halin yanzu - waɗanne mararraba ne muke fuskanta? Ta yaya kwakwalwa da juyin halitta suke aiki, kuma me yasa muke yin amfani da abin da muka riga muka sani game da ci gaban tsarin halitta?
Ci gaban masana'antu na ML ya zo daidai da ci gaban ci gaban Google, wanda ke buga bincikensa akan NeurIPS kowace shekara:
- 1997 - ƙaddamar da wuraren bincike, sabobin farko, ƙananan ikon sarrafa kwamfuta
- 2010 - Jeff Dean ya ƙaddamar da aikin Google Brain, haɓakar hanyoyin sadarwa a farkon farkon.
- 2015 - aiwatar da masana'antu na cibiyoyin sadarwa na jijiyoyi, saurin fahimtar fuska kai tsaye akan na'urar gida, ƙananan matakan sarrafawa waɗanda aka keɓance don ƙididdigar tensor - TPU. Google ya ƙaddamar da Coral ai - analog na rasberi pi, ƙaramin kwamfuta don ƙaddamar da hanyoyin sadarwa na jijiyoyi zuwa na'urorin gwaji.
- 2017 - Google ya fara haɓaka horon da ba a san shi ba tare da haɗa sakamakon horon hanyar sadarwar jijiyoyi daga na'urori daban-daban zuwa samfuri ɗaya - akan Android
A yau, an keɓe gabaɗayan masana'antu don tsaron bayanai, tarawa, da maimaita sakamakon koyo akan na'urorin gida.
- shugabanci na ML wanda samfuran mutum ɗaya ke koya ba tare da juna ba sannan a haɗa su cikin ƙira ɗaya (ba tare da daidaita bayanan tushen ba), an daidaita su don abubuwan da ba a saba gani ba, abubuwan da ba su da kyau, keɓancewa, da sauransu. Duk na'urorin Android da gaske babban kwamfutoci ne guda ɗaya don Google.
Samfuran ƙirƙira bisa haɗin gwiwar ilmantarwa hanya ce mai ban sha'awa a nan gaba a cewar Google, wanda shine "a farkon matakan girma." GANs, a cewar malamin, suna da ikon koyan sake haifar da yawan ɗabi'ar yawan rayayyun halittu da algorithms tunani.
Yin amfani da misalin gine-ginen GAN guda biyu masu sauƙi, an nuna cewa a cikin su neman hanyar ingantawa yana yawo a cikin da'irar, wanda ke nufin ingantawa kamar haka ba ya faruwa. Haka kuma, wa] annan nau'o'in suna da nasara sosai wajen kwaikwayon gwaje-gwajen da masana kimiyyar halittu ke yi a kan yawan kwayoyin cuta, wanda ya tilasta musu su koyi sababbin dabarun halayya don neman abinci. Za mu iya ƙarasa cewa rayuwa tana aiki daban fiye da aikin ingantawa.
Inganta GAN Tafiya
Duk abin da muke yi a cikin tsarin ilmantarwa na inji yanzu ayyuka ne kunkuntar kuma an tsara su sosai, yayin da waɗannan ka'idodin ba su cika da kyau ba kuma ba su dace da ilimin da muke magana a kai a fannoni kamar neurophysiology da ilmin halitta ba.
Abin da ya dace da gaske a lamuni daga fannin neurophysiology a nan gaba shi ne sabbin gine-ginen neuron da ɗan bita kan hanyoyin yada kurakurai.
Ita kanta kwakwalwar dan adam ba ta koyo kamar cibiyar sadarwa ta jijiyoyi:
- Ba shi da abubuwan shigar farko na bazuwar, gami da waɗanda aka shimfida ta hankula da kuma lokacin ƙuruciya
- Yana da inherent kwatance na ci gaban ilhami (sha'awar koyon harshe daga jariri, tafiya a tsaye)
Horar da kwakwalwar mutum aiki ne mara nauyi, watakila ya kamata mu yi la'akari da "mallaka" na mutane masu saurin canzawa da ke ba da ilimi ga juna don sake haifar da hanyoyin juyin halitta.
Abin da za mu iya ɗauka cikin ML algorithms yanzu:
- Aiwatar da ƙirar layin salula waɗanda ke tabbatar da koyo na yawan jama'a, amma gajeriyar rayuwar mutum ("kwakwalwar mutum ɗaya")
- Koyon harbi kaɗan ta amfani da ƙaramin adadin misalai
- Ƙarin hadaddun tsarin neuron, ayyukan kunnawa daban-daban
- Canja wurin "genome" zuwa tsararraki masu zuwa - backpropagation algorithm
- Da zarar mun haɗa neurophysiology da cibiyoyin sadarwa na jijiyoyi, za mu koyi gina kwakwalwa mai aiki da yawa daga sassa da yawa.
Daga wannan ra'ayi, aikin SOTA mafita yana da lahani kuma ya kamata a sake duba shi don bunkasa ayyuka na gama gari (ma'auni).
"Kimiyyar Bayanai ta Gaskiya", Bin Yu (Berkeley)
Rahoton ya keɓe kan matsalar fassarar ƙirar injuna da kuma hanyoyin gwajin su kai tsaye da tabbatarwa. Ana iya fahimtar kowane samfurin ML da aka horar da shi azaman tushen ilimin da ake buƙatar ciro daga gare ta.
A fagage da yawa, musamman a fannin likitanci, yin amfani da abin ƙira ba zai yiwu ba ba tare da fitar da wannan boyayyen ilimi da fassara sakamakon ƙirar ba - in ba haka ba ba za mu tabbatar da cewa sakamakon zai kasance tsayayye, ba bazuwar, abin dogaro, kuma ba zai kashe mai haƙuri. Gabaɗayan tsarin dabarun aiki yana tasowa a cikin tsarin ilmantarwa mai zurfi kuma ya wuce iyakokinsa - kimiyyar bayanai ta gaskiya. Menene shi?
Muna son cimma irin ingancin wallafe-wallafen kimiyya da sake fasalin samfura kamar haka:
- wanda ake iya faɗi
- m
- barga
Waɗannan ka'idoji guda uku sun zama tushen sabuwar hanyar. Ta yaya za a iya bincika ƙirar ML akan waɗannan sharuɗɗan? Hanya mafi sauki ita ce gina samfuran da za a iya fassara nan da nan (regressions, bishiyar yanke shawara). Koyaya, muna kuma son samun fa'idodin zurfafa ilmantarwa nan da nan.
Hanyoyi da yawa na yanzu don aiki tare da matsalar:
- fassara samfurin;
- amfani da hanyoyin bisa hankali;
- yi amfani da tarin algorithms lokacin horarwa, kuma tabbatar da cewa samfuran masu fassarar layi suna koyan tsinkaya amsoshi iri ɗaya kamar hanyar sadarwar jijiyoyi, fassarar fasali daga ƙirar madaidaiciya;
- canji da haɓaka bayanan horo. Wannan ya haɗa da ƙara ƙara, tsangwama, da ƙara bayanai;
- duk wani hanyoyin da ke taimakawa tabbatar da cewa sakamakon samfurin ba bazuwar ba ne kuma ba ya dogara da ƙananan tsangwama maras so (hare-haren abokan gaba);
- fassara samfurin bayan gaskiya, bayan horo;
- nazari yana nuna ma'aunin nauyi ta hanyoyi daban-daban;
- yi nazarin yiwuwar duk hasashe, rarraba aji.
Harin makiya
Kuskuren ƙira suna da tsada ga kowa: babban misali shine aikin Reinhart da Rogov."" ya rinjayi manufofin tattalin arziki na kasashen Turai da dama tare da tilasta musu bin manufofin tsuke bakin aljihu, amma sake duba bayanan da aka yi a hankali da kuma sarrafa su bayan shekaru ya nuna akasin sakamakon!
Duk wata fasaha ta ML tana da tsarin rayuwarta daga aiwatarwa zuwa aiwatarwa. Manufar sabuwar hanyar ita ce bincika ƙa'idodi guda uku a kowane mataki na rayuwar samfurin.
Sakamako:
- Ana haɓaka ayyuka da yawa waɗanda zasu taimaka samfurin ML ya zama abin dogaro. Wannan shi ne, alal misali, zurfafawa (hanyar haɗi zuwa: );
- Don ƙarin haɓaka hanyoyin, yana da mahimmanci don haɓaka ingancin wallafe-wallafe a fagen ML;
- Koyon na'ura yana buƙatar shugabanni masu horo da ƙwarewa a fannonin fasaha da na ɗan adam.
"Tsarin Halin Dan Adam tare da Koyan Injin: Dama da Kalubale" Nuria M Oliver, Albert Ali Salah
Lacca da aka sadaukar don yin kwaikwayon halayen ɗan adam, tushen fasahar sa da kuma abubuwan da ake buƙata.
Za'a iya rarraba ƙirar halayen ɗan adam zuwa:
- mutum hali
- halin ƴan ƙaramin rukuni na mutane
- hali na taro
Kowane ɗayan waɗannan nau'ikan ana iya yin su ta amfani da ML, amma tare da bayanan shigarwa daban-daban da fasali. Kowane nau'i kuma yana da nasa al'amuran da'a waɗanda kowane aikin ke gudana:
- hali na mutum - sata na ainihi, zurfin karya;
- hali na ƙungiyoyin mutane - de-anonymization, samun bayanai game da motsi, kiran tarho, da dai sauransu;
mutum hali
Galibi yana da alaƙa da batun hangen nesa na Computer - fahimtar motsin zuciyar ɗan adam da halayen. Wataƙila kawai a cikin mahallin, a cikin lokaci, ko tare da ma'auni na dangi na bambancin motsin zuciyarsa. Zane-zanen yana nuna fahimtar motsin zuciyar Mona Lisa ta amfani da mahallin daga yanayin tunanin matan Bahar Rum. Sakamako: murmushin farin ciki, amma tare da raini da kyama. Dalili shine mafi kusantar a cikin hanyar fasaha na ayyana motsin "tsaka-tsaki".
Halin ƙaramin rukuni na mutane
Ya zuwa yanzu mafi munin samfurin shine saboda rashin isassun bayanai. Misali, an nuna ayyukan daga 2018 - 2019. akan mutane da yawa X dozin na bidiyo (cf. 100k++ bayanan bayanan hoto). Don mafi kyawun ƙirar wannan ɗawainiya, ana buƙatar bayanin multimodal, zai fi dacewa daga na'urori masu auna firikwensin akan altimeter na jiki, ma'aunin zafi da sanyio, rikodin makirufo, da sauransu.
Halin taro
Yankin da ya fi ci gaba, tun da abokin ciniki shine Majalisar Dinkin Duniya da jihohi da yawa. Kyamarar sa ido na waje, bayanai daga hasumiya na tarho - lissafin kuɗi, SMS, kira, bayanai akan motsi tsakanin iyakokin jihohi - duk wannan yana ba da hoto mai inganci na motsin mutane da rashin zaman lafiya. Mahimman aikace-aikace na fasaha: inganta ayyukan ceto, taimako da ƙaurawar jama'a a lokacin gaggawa. Samfuran da aka yi amfani da su galibi har yanzu ba a fassara su ba - waɗannan su ne LSTMs iri-iri da hanyoyin sadarwa na juyin juya hali. Akwai taƙaitaccen bayani cewa Majalisar Ɗinkin Duniya tana fafutukar ganin an samar da sabuwar doka da za ta tilastawa 'yan kasuwan Turai su raba bayanan da ba a bayyana sunansu ba don kowane bincike.
"Daga Tsarin 1 zuwa Tsarin 2 Zurfafa Koyo", Yoshua Bengio
A cikin laccar Joshua Bengio, zurfafa ilmantarwa ta haɗu da ilimin halin ɗan adam a matakin saitin manufa.
Bengio ya gano manyan nau'ikan matsaloli guda biyu bisa ga dabarar wanda ya lashe kyautar Nobel Daniel Kahneman (littafin "")
nau'in 1 - Tsarin 1, ayyukan da ba a sani ba da muke yi "ta atomatik" (kwakwalwar tsohuwar): tuki mota a wuraren da aka saba, tafiya, gane fuskoki.
nau'in 2 - Tsarin 2, ayyuka masu hankali (cerebral cortex), saitin manufa, bincike, tunani, ayyuka masu haɗaka.
AI ya zuwa yanzu ya kai isassun madaidaici kawai a cikin ayyuka na nau'in farko, yayin da aikinmu shine kawo shi zuwa na biyu, muna koyar da shi don aiwatar da ayyuka da yawa da aiki tare da dabaru da ƙwarewar fahimi.
Don cimma wannan burin an tsara shi:
- a cikin ayyukan NLP, yi amfani da hankali azaman maɓalli mai mahimmanci don ƙirar tunani
- yi amfani da koyo-koyo da wakilcin koyo don ingantattun fasalulluka na ƙira waɗanda ke yin tasiri ga wayewa da kasancewarsu - kuma a kan tushensu sun ci gaba zuwa aiki tare da manyan ra'ayoyi.
Maimakon kammalawa, ga jawabin da aka gayyata: Bengio yana ɗaya daga cikin masana kimiyya da yawa waɗanda ke ƙoƙarin faɗaɗa fannin ML fiye da matsalolin ingantawa, SOTA da sababbin gine-gine.
Tambayar ta kasance a buɗe har zuwa wane nau'i na haɗuwa da matsalolin sani, tasirin harshe akan tunani, neurobiology da algorithms shine abin da ke jiran mu a nan gaba kuma zai ba mu damar matsawa zuwa na'urori masu "tunani" kamar mutane.
Na gode!
source: www.habr.com