NeuroIPS () ndiwo musangano mukuru pasi rose wekudzidza muchina uye hungwaru hwekugadzira uye chiitiko chikuru munyika yekudzidza kwakadzama.
Isu, mainjiniya eDS, tichavewo nyanzvi biology, linguistics, uye psychology mumakore gumi matsva? Tichakuudza muongororo yedu.
Gore rino musangano uyu wakaunganidza vanhu vanodarika zviuru gumi nezvitatu nemazana mashanu kubva kunyika makumi masere muVancouver, Canada. Iri harisi gore rokutanga iro Sberbank yakamiririra Russia pamusangano - boka reDS rakataura nezvekushandiswa kweML mumabhangi maitiro, pamusoro pemakwikwi eML uye pamusoro pekugona kweSberbank DS platform. Ndeapi aive maitiro makuru e13500 munharaunda yeML? Vatori vemusangano vanoti: ΠΈ .
Gore rino, NeurIPS yakagamuchira anopfuura 1400 mapepa-algorithms, mhando nyowani, uye mitsva yekushandisa kune data nyowani.
Zviri Mukati:
- Trends
-
- Muenzaniso kududzira
- Multidisciplinarity
- Kukurukurirana
- RL
- GAN
- Hurukuro Dzakakokwa Dzinokosha
-
- "Social Intelligence", Blaise Aguera naArcas (Google)
- "Veridical Data Sayenzi", Bin Yu (Berkeley)
- "Human Behaviour Modelling ne Machine Kudzidza: Mikana uye Matambudziko", Nuria M Oliver, Albert Ali Salah
- "Kubva kuSystem 1 kusvika kuSystem 2 Kudzidza Kwakadzika", Yoshua Bengio
Trends 2019
1. Kududzira kwemuenzaniso uye nzira itsva yeML
Musoro mukuru wemusangano kududzira uye humbowo hwekuti sei tichiwana mimwe mibairo. Mumwe anogona kutaura kwenguva yakareba pamusoro pekukosha kwehuzivi hwekududzirwa kwe "black box", asi pakanga paine dzimwe nzira dzechokwadi uye hunyanzvi hwekuvandudza munzvimbo ino.
Nzira yekudzokorora modhi uye kutora ruzivo kubva mazviri ibhuku idzva rekushandisa resainzi. Mamodheru anogona kushanda sechishandiso chekuwana ruzivo rutsva nekuchiyedza, uye nhanho imwe neimwe yepreprocessing, kudzidziswa uye kushandiswa kweiyo modhi inofanirwa kudhindwa.
Chikamu chakakosha chezvinyorwa zvakapirwa kwete mukuvakwa kwemamodheru uye maturusi, asi kune matambudziko ekuvimbisa kuchengetedzwa, pachena uye kuvimbiswa kwemhedzisiro. Kunyanya, rukova rwakaparadzana rwakaonekwa nezve kurwiswa kwemuenzaniso (kurwiswa nevavengi), uye sarudzo dzekurwiswa kwese kurwiswa uye kurwiswa kwekushandisa kunotariswa.
Zvinyorwa:
- - chinyorwa chechirongwa chemaitiro ekusimbisa muenzaniso. Inosanganisira tarisiro yezvishandiso zvemazuva ano zvekududzira modhi, kunyanya, kushandiswa kwekutarisisa uye kuwana kukosha kwechimiro ne "distilling" iyo neural network ine mitsetse mhando.
- 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 inoratidza kududzira modhi yemabasa ekugadzirisa zvinyorwa
2. Multidisciplinarity
Kuti tive nechokwadi chechokwadi chechokwadi uye nekugadzira nzira dzekusimbisa nekuwedzera ruzivo, tinoda nyanzvi mundima dzine hukama avo panguva imwe chete vane hunyanzvi muML uye munzvimbo yezvidzidzo (mushonga, linguistics, neurobiology, dzidzo, nezvimwewo). Zvinonyanya kukosha kucherechedza kuvepo kwakakosha kwemabasa nekutaura mune neuroscience uye cognitive sainzi - kune kuwirirana kwenyanzvi uye kukwereta kwemazano.
Pamusoro peiyo rapprochement, multidisciplinarity iri kubuda mukubatanidzwa kweruzivo kubva kwakasiyana zvinyorwa: zvinyorwa nemifananidzo, zvinyorwa nemitambo, girafu dhatabhesi + zvinyorwa uye mapikicha.
Zvinyorwa:
- Neuroscience + ML -
- VisualQA -
- RL + NLP -
Mhando mbiri - Strategist uye Executive - yakavakirwa paRL uye NLP tamba online zano
3. Kukurukurirana
Kusimbisa hungwaru hwekugadzira idanho rakanangana nekuzvidzidzira masisitimu, "kuziva", kufunga uye kufunga. Kunyanya, causal inference uye commonsense kufunga kuri kukura. Mimwe mishumo yakazvipira kune meta-kudzidza (zvekuti ungadzidza sei kudzidza) uye musanganiswa weDL matekinoroji ane 1st uye 2nd order logic - izwi rekuti Artificial General Intelligence (AGI) riri kuita izwi rinozivikanwa mukutaura kwevatauri.
Zvinyorwa:
- 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.Kusimbisa Kudzidza
Mazhinji ebasa anoenderera mberi nekuvandudza nzvimbo dzechinyakare dzeRL - DOTA2, Starcraft, kubatanidza zvivakwa zvine komputa kuona, NLP, graph dhatabhesi.
Zuva rakasiyana remusangano rakatsaurirwa kumusangano weRL, apo iyo Optimistic Actor Critic Model architecture yakaunzwa, yepamusoro kune ese apfuura, kunyanya Soft Actor Critic.
Zvinyorwa:
- ; Kamil Ciosek, Quan Vuong, Robert Loftin, Katja Hofmann
- ; Yasuhiro Fujita (Preferred Networks, Inc.)*; Toshiki Kataoka (Preferred Networks, Inc.); Prabhat Nagarajan (Inodiwa Networks); Takahiro Ishikawa (The University of Tokyo) [external pdf link].
- ; Danijar Hafner (Google)*; Timothy Lillicrap (DeepMind); Jimmy Ba (Yunivhesiti yeToronto); Mohammad Norouzi (Google Brain)
Vatambi veStarCraft vanorwa neAlphastar modhi (DeepMind)
5.GAN
Manetiweki ekugadzira achiri mukutarisa: mabasa mazhinji anoshandisa vanilla GANs ehumbowo hwemasvomhu, uye zvakare anoashandisa munzira nyowani, dzisina kujairika (magirafu ekugadzira modhi, kushanda nenhevedzano, kunyorera kukonzeresa-uye-mhedzisiro hukama mune data, nezvimwewo).
Zvinyorwa:
- Sangwoo Mo, Chiheon Kim, Sungwoong Kim, Minsu Cho, Jinwoo Shin
- Dan Zhang, Anna Khoreva
- Lei Xu, Maria Skoularidou, Alfredo Cuesta-Infante, Kalyan Veeramachaneni
Sezvo basa rakawanda rakagamuchirwa Pasi apa tichataura pamusoro pehurukuro dzinonyanya kukosha.
Hurukuro Dzakakokwa
"Social Intelligence", Blaise Aguera naArcas (Google)
Hurukuro iyi inotarisa pane yakajairika nzira yekudzidza muchina uye tarisiro yekuchinja indasitiri izvozvi - ndeipi mharadzano yatakatarisana nayo? Uropi nemhindumupindu zvinoshanda sei, uye nei tisingashandisi zvishoma zvatinoziva nezvekugadzirwa kwezvinhu zvakasikwa?
Kuvandudzwa kwemaindasitiri kweML kunopindirana zvakanyanya nehukuru hwekuvandudzwa kweGoogle, iyo inoburitsa tsvakiridzo yayo paNeurIPS gore negore:
- 1997 - kutangwa kwenzvimbo dzekutsvaga, yekutanga maseva, diki komputa simba
- 2010 - Jeff Dean anotanga chirongwa cheGoogle Brain, iyo boom yeneural network pakutanga chaipo.
- 2015 - kushandiswa kwemaindasitiri kweneural network, kukurumidza kuzivikanwa kwechiso zvakananga pane yemuno mudziyo, yakaderera-level processors yakagadzirirwa tensor komputa - TPU. Google inovhura Coral ai - analogue yeraspberry pi, mini-kombuta yekuunza neural network mukuisirwa kwekuyedza.
- 2017 - Google inotanga kusimudzira kudzidziswa kwakadzikama uye kusanganisa mhedzisiro yeneural network kudzidziswa kubva kune akasiyana michina kuita modhi imwe - pa Android.
Nhasi, indasitiri yese yakatsaurirwa kuchengetedza data, kuunganidza, uye kudzokorodza kwezvinobuda pakudzidza pamidziyo yemuno.
-Nhungamiro yeML umo mamodheru ega ega anodzidza akazvimirira kubva kune mumwe nemumwe uye obva asanganiswa kuita imwe modhi (pasina kuisa pakati peiyo sosi data), yakagadziridzwa zviitiko zvisingawanzo, anomalies, personalization, nezvimwe. Ese maturusi eAroid anonyanya kuita imwe komputa huru yeGoogle.
Mamodheru ekugadzira akavakirwa pakudzidza kwakabatana igwara rinovimbisa remangwana maererano neGoogle, iri "mumatanho ekutanga ekukura kukuru." MaGAN, maererano nemudzidzisi, anokwanisa kudzidza kubereka maitiro akawanda ehuwandu hwezvipenyu uye kufunga algorithms.
Uchishandisa muenzaniso wemagadzirirwo maviri akareruka eGAN, zvinoratidzwa kuti mavari kutsvaga kwenzira yekuvandudza kunodzungaira mudenderedzwa, izvo zvinoreva kuti optimization yakadaro haiitike. Panguva imwecheteyo, mienzaniso iyi inobudirira zvikuru mukuenzanisa zviedzo izvo biologists vanoita pahuwandu hwebhakitiriya, vachivamanikidza kudzidza maitiro matsva ekutsvaga kwekudya. Tinogona kugumisa kuti hupenyu hunoshanda zvakasiyana pane optimization basa.
Kufamba GAN Optimization
Zvese zvatinoita muchimiro chekudzidza muchina izvozvi mabasa akamanikana uye akanyanya kurongeka, nepo maformalism aya asingaite zvakanaka uye haaenderane neruzivo rwechidzidzo chedu munzvimbo dzakaita seneurophysiology uye biology.
Chinonyanya kukosha kukwereta kubva kumunda weurourophysiology munguva pfupi iri kutevera ndeye neuron architectures uye kudzokororwa kudiki kwemaitiro ekudzosera kumashure zvikanganiso.
Huropi hwemunhu pachahwo hahudzidzi senge neural network:
- Haana mapindiro ekutanga ekutanga, kusanganisira ayo akaiswa pasi kuburikidza nepfungwa uye muhudiki
- Ane maitirwo emuzvarirwo ekukura kwemusikirwo (chido chokudzidza mutauro kubva mucheche, kufamba wakatwasanuka)
Kudzidzisa huropi hwemunhu ibasa rakaderera; pamwe isu tinofanirwa kufunga nezve "makoloni" evanhu vari kukurumidza kushandura vachipfuudza ruzivo kune mumwe nemumwe kuti vabereke nzira dzekushanduka kweboka.
Zvatinogona kutora muML algorithms izvozvi:
- Isa mamodhi emutsara wemaseru anovimbisa kudzidza kwehuwandu, asi hupenyu hupfupi hwemunhu ("huropi hwemunhu")
- Vashoma-kupfura kudzidza uchishandisa nhamba shoma yemuenzaniso
- Zvimwe zvakaoma neuron zvimiro, zvishoma zvakasiyana activation mabasa
- Kuendesa iyo "genome" kuzvizvarwa zvinotevera - backpropagation algorithm
- Kana tangobatanidza neurophysiology uye neural network, isu tichadzidza kuvaka huropi huzhinji hunoshanda kubva kuzvinhu zvakawanda.
Kubva pane iyi maonero, maitiro eSOTA mhinduro anokanganisa uye anofanirwa kudzokororwa nekuda kwekugadzira mabasa akafanana (benchmarks).
"Veridical Data Sayenzi", Bin Yu (Berkeley)
Chirevo chakazvipira kune dambudziko rekududzira modhi yekudzidza yemuchina uye nzira yekuyedza kwavo kwakananga nekusimbisa. Chero yakadzidziswa ML modhi inogona kunzwisiswa sesosi yeruzivo inoda kutorwa kubva mairi.
Munzvimbo dzakawanda, kunyanya mukurapa, kushandiswa kwemuenzaniso hakugoneki pasina kubvisa ruzivo urwu rwakavanzika uye kududzira migumisiro yemuenzaniso - kana zvisina kudaro isu hatizove nechokwadi chokuti migumisiro ichave yakagadzikana, isiri-random, yakavimbika, uye haizourayi murwere. Iyo nzira yese yemaitiro ebasa iri kukura mukati meiyo yakadzika yekudzidza paradigm uye inodarika miganhu yayo - veridical data sainzi. Chii?
Tinoda kuwana mhando yakadai yezvinyorwa zvesainzi uye kugadzirwazve kwemamodheru ayo ari:
- zvinofanotaurwa
- computable
- stable
Misimboti mitatu iyi inoumba hwaro hwenzira itsva. Maitiro eML angatariswe sei achipesana nemaitiro aya? Nzira iri nyore ndeyekuvaka nekukurumidza kududzira modhi (regressions, sarudzo miti). Zvisinei, tinodawo kuwana mabhenefiti epakarepo ekudzidza kwakadzama.
Nzira dzinoverengeka dziripo dzekushanda nedambudziko:
- kududzira muenzaniso;
- shandisa nzira dzinobva pakutarisisa;
- shandisa ensembles yealgorithms paunenge uchidzidzira, uye ive nechokwadi chekuti mamodheru anodudzirwa anodzidza kufanotaura mhinduro dzakafanana neneural network, kududzira maficha kubva kumutsara modhi;
- kuchinja uye kuwedzera kudzidziswa data. Izvi zvinosanganisira kuwedzera ruzha, kupindira, uye kuwedzera data;
- chero nzira dzinobatsira kuvimbisa kuti mhedzisiro yemuenzaniso haina kurongeka uye haina kutsamira pane kudiki kupindira kusingadiwi (kurwiswa kweanopikisa);
- kududzira muenzaniso mushure mechokwadi, mushure mekudzidziswa;
- kudzidza kunoratidza zviyero zvakasiyana-siyana;
- dzidza zvingangoitika zvekufungidzira kwese, kugovera kwekirasi.
Adversarial attack
Zvikanganiso zvekuenzanisira zvinodhura kumunhu wese: muenzaniso wekutanga ibasa raReinhart naRogov." yakapesvedzera marongero ehupfumi hwenyika zhinji dzeEurope uye ikavamanikidza kuti vatevedze marongero akashata, asi kunyatsoongororazve data uye kugadzirisa kwavo makore gare gare kwakaratidza mhedzisiro!
Chero tekinoroji yeML ine kutenderera kwayo kwehupenyu kubva pakuitwa kusvika pakuitwa. Chinangwa cheiyo nzira nyowani ndeyekutarisa pamisimboti mitatu yakakosha padanho rega rega rehupenyu hwemuenzaniso.
Mhinduro:
- Mapurojekiti akati wandei ari kugadzirwa anozobatsira iyo ML modhi kuva yakavimbika. Izvi, semuenzaniso, deeptune (link ku: );
- Kuti uwedzere kuvandudzwa kweiyo nzira, zvinodikanwa kuvandudza zvakanyanya kunaka kwezvinyorwa mumunda weML;
- Kudzidza kwemuchina kunoda vatungamiriri vane dzidziso dzakasiyana siyana uye hunyanzvi mune zvese zvehunyanzvi uye zvevanhu.
"Human Behaviour Modelling ne Machine Kudzidza: Mikana uye Matambudziko" Nuria M Oliver, Albert Ali Salah
Chidzidzo chakatsaurirwa kuenzanisira maitiro evanhu, nheyo dzayo dzetekinoroji uye tarisiro yekushandisa.
Maitiro evanhu anogona kukamurwa kuva:
- hunhu hwemunhu
- maitiro eboka duku revanhu
- maitiro akawanda
Imwe neimwe yemhando idzi inogona kuenzanisirwa uchishandisa ML, asi iine ruzivo rwakasiyana rwekuisa uye maficha. Rudzi rwega rwega runewo nyaya dzayo dzetsika idzo purojekiti yega yega inofamba nayo:
- hunhu hwemunhu - kubiwa kwekuzivikanwa, deepfake;
- maitiro emapoka evanhu - de-anonymization, kuwana ruzivo nezve mafambiro, mafoni efoni, nezvimwewo;
hunhu hwemunhu
Zvikuru zvine hukama nemusoro weComputer Vision - kucherechedzwa kwemanzwiro evanhu uye maitiro. Zvichida chete mumamiriro ezvinhu, nekufamba kwenguva, kana nemwero wepakati wekusiyana kwake kwemanzwiro. Iyo siraidhi inoratidza kucherechedzwa kwemanzwiro aMona Lisa uchishandisa mamiriro kubva kumanzwiro ekuti vakadzi veMedithera. Mhedzisiro: kunyemwerera kwemufaro, asi nekuzvidza uye kusemeswa. Chikonzero chinogona kunge chiri munzira yehunyanzvi yekutsanangura manzwiro e "kusarerekera".
Maitiro eboka diki revanhu
Parizvino iyo yakaipisisa modhi inokonzerwa neruzivo rusina kukwana. Semuenzaniso, mabasa kubva ku2018 - 2019 akaratidzwa. pavanhu vakawanda X madhazeni emavhidhiyo (cf. 100k++ mifananidzo yedatasets). Kuti uenzanise zvakanakisa basa iri, ruzivo rwemultimodal runodiwa, zviri nani kubva kuma sensors pane altimeter yemuviri, thermometer, kurekodha maikorofoni, nezvimwe.
Kuzvibata kwevanhu vakawanda
Iyo yakanyanya kusimukira nzvimbo, sezvo mutengi ari UN uye nyika zhinji. Kunze kwekutarisa makamera, data kubva kunhare dzenhare - kubhadharisa, SMS, mafoni, data pamusoro pekufamba pakati pemiganhu yenyika - zvese izvi zvinopa mufananidzo wakavimbika kwazvo wekufamba kwevanhu uye kusagadzikana munharaunda. Zvingangove zvekushandisa tekinoroji: optimization yemabasa ekununura, rubatsiro uye kuburitswa kwehuwandu panguva yekukurumidza. Mamodheru anoshandiswa haasati anyatso kududzirwa - aya akasiyana maLSTM uye convolutional network. Paive nekutaura muchidimbu kuti UN yaitsvaga mutemo mutsva waizomanikidza mabhizinesi eEurope kugovana data risingazivikanwe rinodiwa kune chero tsvagiridzo.
"Kubva kuSystem 1 kusvika kuSystem 2 Kudzidza Kwakadzika", Yoshua Bengio
Muhurukuro yaJoshua Bengio, kudzidza kwakadzama kunosangana neuroscience padanho rekumisa chinangwa.
Bengio anozivisa marudzi maviri makuru ematambudziko zvinoenderana nemaitiro emubairo weNobel Daniel Kahneman (bhuku "")
mhando 1 - System 1, zviito zvisingazivi zvatinoita "otomatiki" (uropi hwekare): kutyaira motokari munzvimbo dzakajairika, kufamba, kuziva zviso.
mhando 2 - System 2, zviito zvekuziva (cerebral cortex), kuisa chinangwa, kuongorora, kufunga, mabasa akaumbwa.
AI kusvika pari zvino yasvika pakakwirira zvakakwana chete mumabasa erudzi rwekutanga, nepo basa redu nderekuunza kune yechipiri, tichiidzidzisa kuita mabasa mazhinji uye kushanda zvine musoro uye nepamusoro-soro hunyanzvi hwekuziva.
Kuzadzisa chinangwa ichi zvinokurudzirwa:
- muNLP mabasa, shandisa kutarisisa senzira yakakosha yekuenzanisira kufunga
- shandisa meta-kudzidza uye kumiririra kudzidza kune zvirinani modhi maficha anopesvedzera kuziva uye nharaunda yavo - uye pahwaro hwavo vanoenderera mberi nekushanda nepamusoro-level concepts.
Panzvimbo yekupedzisa, heino hurukuro yakakokwa: Bengio ndomumwe wevazhinji vesainzi vari kuyedza kuwedzera ndima yeML kupfuura matambudziko ekugadzirisa, SOTA uye zvivakwa zvitsva.
Mubvunzo unoramba wakavhurika kusvika papi kusanganiswa kwematambudziko ekuziva, pesvedzero yemutauro pakufunga, neurobiology uye algorithms ndizvo zvakatimirira mune ramangwana uye zvichatibvumidza kutamira kumichina "inofunga" sevanhu.
Ndinokutendai!
Source: www.habr.com