Kwenzekeni
Google Onjiniyela bashicilele umchazi wezobuchwepheshe ngo-Agasti 27 ochaza ukuthi abacwaningi e-Pierre Auger Observatory basebenzise kanjani inethiwekhi ye-neural esekelwe e-Keras ukuze bahlaziye izihlambi zomoya ze-cosmic-ray. Okuthunyelwe kwethula imiphumela yesayensi njengomsebenzi oshicilelwe ngaphambilini, hhayi njengesilinganiso esisanda kumenyezelwa.
Okuthunyelwe kuchaza ukuthi izinhlayiya ze-cosmic ezinamandla aphezulu ngokuvamile azitholakali ngokuqondile. Lapho zingena emkhathini womhlaba, zikhiqiza izihlambi zomoya ezibanzi: ama-cascade ezinhlayiya zesibili ezisakazeka ezindaweni ezinkulu. Ama-Observatory afana ne-Pierre Auger Observatory ngakho-ke zisebenzisa iziteshi zomtshina ezisatshalaliswe kabanzi ukuze isampula indawo yeshawa futhi irekhode imikhondo yesignali ye-nanosecond-resolution. Abacwaningi kumele bacabangele uhlobo lwezinhlayiyana zangempela, amandla kanye nesiqondiso kusuka ezilinganisweni ezingaphelele nezingaqondile. Umthombo uchaza lokhu njengenkinga ephambene elawulwa ukungahleliwe okubonakalayo, ngoba ngisho nezihlambi ezikhiqizwa izinhlayiya ezifanayo ezingenayo ngeke zikhule ngendlela efanayo ncamashi.
I-Architecture ye-Keras echazwe kokuthunyelwe yakhelwe eduze kwaleso sakhiwo. Ngedatha ye-Auger, imodeli ithatha i-cutout yesiteshi engu-13-by-13 egxile esiteshini esinesiginali enkulu kunazo zonke. Isiteshi ngasinye sinikela ngemikhondo emithathu ye-waveform, ngasinye siqukethe izinyathelo zesikhathi ezingu-120 ngezikhawu ezingama-25-nanosecond, kanye nenani lesikhathi sokufika kanye nenkomba yesimo ebonisa ukuthi isiteshi siyasebenza noma silahlekile. Isifaki khodi sesikhashana sisebenza izendlalelo ze-LSTM eziqondiswe kabili nezijwayelekile ekulandeleleni isiteshi ngasinye. Ngenxa yokuthi inethiwekhi engaphansi efanayo yesikhashana iphinda isetshenziswe kugridi yonkana, imodeli ithatha ngokuthi i-physics yomtshina efanelekile iyahambisana ukusuka kwesinye isiteshi ukuya kwesinye.
Izici zesikhashana zibe sezihlanganiswa nesikhathi nolwazi lwesimo bese zidluliselwa kunethiwekhi yendawo. Okuthunyelwe kuthi le ngxenye isebenzisa ama-hexagonal, ama-convolutions alingana neqembu ukuze abonise ukwakheka komtshina we-observatory kanye nokulingana okujikelezayo kwezihlambi zomoya. Ibhulokhi exhumeke kakhulu igcina izici zangaphambili, okulandelwa ukucutshungulwa okuqondene nomsebenzi othile wamandla, ukujula kobukhulu beshawa, ingqikithi yeshawa kanye nesiqondiso sokufika. Umthombo uthi inethiwekhi yaqeqeshwa kusukela ekupheleni kumasignali omtshina alingisa. Yethula umphumela wesayensi obikiwe njengesifinyezo somsebenzi oshicilelwe ku-JINST kanye Nezincwadi Zokubukeza Umzimba, kunobufakazi bomphumela omusha okhiqizwe yile bhulogi.
Imininingwane yomthombo: developers.googleblog.com ↗
Kungani kubalulekile
Indlela ikhombisa ukuthi i-AI ingalukhipha kanjani ulwazi kumafomu omtshina omtshina agcweleyo lawo izindlela ezijwayelekileyo ezicindezela ekutshayeleni kanye lesikhathi sokufika. Ngokomthombo, lokhu kunwebe idathasethi esebenzisekayo futhi kwasiza ukuveza ukuthi imisebe ye-cosmic iba nzima ngamandla amakhulu, nakuba izimangalo zidinga ukulinganisa ngokucophelela ngokumelene nedatha yomtshina wangempela.
Ushintsho oluyinhloko lobuchwepheshe ukuthi inethiwekhi isebenzisa umumo kanye nesikhathi samagagasi aqoshiwe kunokuthembela ngokuyinhloko enanini elicindezelweyo: isignali ehlanganisiwe, noma ukushaja, kanye nesikhathi sokufika kwezinhlayiyana zokuqala. Okuthunyelwe kuthi lokho okokufaka kwendabuko kusekela ukwakhiwa kabusha kwamandla awusizo kodwa akufanelekile kahle ukuhlukanisa imiphumela ecashile yesisindo se-cosmic-ray. I-nucleus esindayo ingadala ama-sub-shower amaningi kanye nengxenye enkulu ye-muon ehlobene, ekhiqiza izici ze-waveform okungase kube nzima ukuzithwebula ngezifinyezo eziklanywe ngesandla. Ukufunda ngokuqondile emikhondweni kunikeza imodeli ukufinyelela kuleso sakhiwo sesikhashana.
Imithelela yesayensi echazwe umthombo ibalulekile uma ukwakhiwa kabusha okubikiwe kuqinisekiswa. Okuthunyelwe kuthi inethiwekhi ivule idathasethi enkulu ngokuphindwe kashumi kunokubonwa kwezibonakude zesimanje, isebenzisa umtshina owawungaklanyelwe ukukala ukwakheka kwe-cosmic-ray. Ithi umphumela ofanayo wokubonwa kwezibonakude ungadinga cishe ikhulu leminyaka yokusebenza okuqhubekayo. Umthombo uqhubeka ubika ukuthi ukuhlaziya kuthole ukuthi imisebe ye-cosmic iba nzima ngokuqhubekayo ngamandla aphezulu kanye nesakhiwo esihlonziwe kuleyo threndi yokuqamba eqondana nezici ezaziwayo ze-spectrum yamandla e-cosmic-ray. Lokho kubuka kungasiza abacwaningi baphenye ukuthi izinhlayiya ezinamandla kakhulu zisheshiswa kuphi futhi kanjani.
Lesi isibonelo se-AI ekhulisa inani elisebenzayo lethuluzi lesayensi elikhona, kunokuthuthukisa ngokomzimba umtshina. Okuthunyelwe kuthi izindlela zokufunda ezijulile ezifanayo e-IceCube zisize ukwakha kabusha nokukhetha imicimbi futhi zaba nomthelela ekutholweni kwe-neutrinos endizeni ye-galactic. Iphinde ithi ukufunda okujulile kungabuyisa izehlakalo ezingabandakanyi amapayipi okwakha kabusha avamile njengezinzima kakhulu ukuzihlaziya, okungase kube ngcono kwezibalo zenhlolovo nokwenza ukuba kulandeleke ngokushesha. Kodwa ukusebenza kokubikezela kwemodeli akufani nokuchazwa ngokomzimba. Umthombo uxwayisa ngokuthi amasistimu ebhokisi elimnyama angahweba ngokutolika ukuze kube nokunemba nokuthi izilinganiso eziqondile zokungaqiniseki zibalulekile uma imiphumela isetshenziselwa ukuhlola okucatshangwayo noma ukuhlola okungase kutholwe.
I-Interactive Mechanism: Indlela Esebenza Ngayo Ngempela
Hlola ubuchwepheshe obuyisisekelo ngemuva kwalokhu kuthuthukiswa ngokuhlanganyela.
Which component of an AI application is the machine-learning model itself?
Ongakubuka ngokulandelayo
Ukuhlola okubalulekile ukuthi amamodeli aqeqeshelwe ezilinganisweni ahlala ethembekile yini ekuhlolweni okuzimele futhi akhiqize izilinganiso ezinembile zokungaqiniseki. Okuthunyelwe kuphinde kukhombe kumanethiwekhi e-neural egrafu, ama-point-cloud transformer kanye namamodeli esisekelo sokuhlola okuphambene njengabangase balandele, kodwa akuhlinzeki ngemiphumela ebonisa ukuthi lezi zinhlelo zesikhathi esizayo zisebenza ngokuzijwayeza.
Inkinga ekhona manje igebe phakathi kokulingisa nokubuka. Inethiwekhi yaqeqeshelwa amasignali omtshina alingisa, kuyilapho amathuluzi angempela anomehluko wokulinganisa, iziteshi ezihlulekile, izifunda ezingekho kanye nezimo okungenzeka zingamelwe ngokuphelele ekulingiseni. Okuthunyelwe kuthi ukulungiswa kwesizinda, ukulinganisa ngezithonjana ezihlukene kanye nokuqinisekiswa okuphambene ngokumelene nedatha ezimele kubalulekile. Iphinde igcizelele ukuthi izilinganiso zokungaqiniseki kufanele zibhekane nokungaqiniseki kwemodeli kanye namashifu phakathi kokusabalalisa okulingiswayo nokwangempela. Umthombo awunikezi imigoqo yephutha yobuningi, imiphumela yokuqinisekisa ezimele noma i-akhawunti enemininingwane yokuthi ukukhuphuka okubikiwe okuphindwe kashumi kukalwe kanjani.
I-architecture ngokwayo nayo ingase ishintshe. Okuthunyelwe kuthi amanethiwekhi we-graph neural kanye nama-point-cloud transformer angase afaneleke kangcono emithonjeni engumtshina engajwayelekile, kuyilapho amamodeli esikhashana asuselwa ku-transformer angangena esikhundleni se-LSTM ebiza kakhulu ngekhompyutha. Lezi zindlela ezihlukile zizodinga ukubonisa ngaphezu kwamaphuzu athuthukisiwe emicimbini elungisiwe. Ubufakazi obuwusizo bungabandakanya ukusebenza kudatha ye-observatory ebanjiwe, ukuqina kuzitholi ezingekho, ukuzinza kuwo wonke ububanzi bamandla kanye nezilinganiso ezithembekile zalapho imodeli ingaqinisekile. Umthombo awunikezi imiphumela enjalo yokuqhathanisa yezakhiwo ezintsha.
Amathuba esikhathi eside imodeli eyisisekelo eqeqeshwe kuzo zonke izivivinyo eziningi nezinhlobo zezithunywa, okuhlanganisa imisebe ye-cosmic, ama-neutrino namagagasi adonsela phansi. Okuthunyelwe kwethula lokhu njengesikhombisi-ndlela esithembisayo sokwakhiwa kabusha komcimbi, ukulingisa nokwambuleka, hhayi njengesistimu ekhona noma ukusetshenziswa. Abacwaningi bazodinga ukuthola ukuthi ukuhlanganisa idatha evela ezindaweni zokubuka ezihlukene kuthuthukisa ukwenziwa okuvamile ngaphandle kokufihla ukuchema okuqondene nensimbi ethile. Bazophinda badinge izindlela ezidalula izindlela eziyimbangela, ama-symmetries nemithetho yemvelo kunokuthuthukisa kuphela ukunemba kokubikezela. Okwamanje, ukuthuthukiswa okucacile kuyi-akhawunti enemininingwane yokuthi ukufunda okujulile sekusetshenziswe kanjani kudatha enzima ye-astroparticle, kanye nesixwayiso esicacile sokuthi ukwethembeka kwesayensi kuncike ekuqinisekiseni nasekuchazeni.