Riaan Isici I-Generative AI ibeka izinselele ezithakazelisayo kubashicileli bezemfundo ababhekana nokukhwabanisa emaphepheni esayensi njengoba ubuchwepheshe bubonisa amandla okukhohlisa ukubuyekezwa kontanga.
Chaza isithombe se-DALL-E, Stable Diffusion, neMidjourney, futhi zizokhiqiza esisodwa ngesekhondi. Lezi zinhlelo zokuguqula umbhalo uye ezithombeni zithuthuke ngokushesha eminyakeni embalwa edlule futhi lokho okwaqala njengesibonelo socwaningo, kwakhiqiza ukulunga nokuxaka ngendlela emangalisayo. imifanekiso yezingane ze-daikon radishes ezihamba izinja ngo-2021, selokhu zashintsha zibe isofthiwe yezohwebo, eyakhiwe izinkampani zamadola ayizigidi eziyizinkulungwane, ezikwazi ukukhiqiza izithombe ezingokoqobo ezikhulayo.
Lawa mamodeli e-AI angakhiqiza izithombe ezinjengokuphila zobuso bomuntu, izinto, nezigcawu, futhi kubukeka njengendaba yesikhathi ngaphambi kokuba zithole kahle ekudaleni izithombe zesayensi nedatha ekholisayo. Amamodeli okuthumela umbhalo-ukuya-isithombe manje asefinyeleleka kabanzi, ashibhile kakhulu ukuwasebenzisa, futhi angasiza ososayensi abangazi kahle ukuthi benze imiphumela futhi bashicilele ucwaningo mbumbulu kalula.
Ukukhohliswa kwesithombe sekuyinto ekhathaza kakhulu abashicileli bezemfundo njengoba kuyikhona kakhulu ifomu evamile yokungaziphathi kahle kwezesayensi kwakamuva. Ababhali bangasebenzisa zonke izinhlobo zamaqhinga, njengokuphequlula, ukuzungezisa, noma ukunqampuna izingxenye zesithombe esifanayo kudatha mbumbulu. Abahleli bakhohliswa ukuthi yonke imiphumela ethulwayo ingeyangempela futhi bazoshicilela umsebenzi wabo.
Abashicileli abaningi manje sebephendukela kusoftware ye-AI ngomzamo wokwenza lokho thola izimpawu zokuphindwaphindwa kwezithombe phakathi nenqubo yokubuyekeza. Ezimweni eziningi, izithombe ziphindwe ngephutha ososayensi abahlanganise idatha yabo, kodwa ngezinye izikhathi zisetshenziselwa ukukhwabanisa okusobala.
Kodwa njengoba abashicileli beqala ukubamba ukuphindwaphindwa kwezithombe, kuvela olunye usongo. Abanye abacwaningi bangase balingeke ukuba basebenzise amamodeli e-AI akhiqizayo ukuze bakhe idatha engelona iqiniso. Eqinisweni, kunobufakazi obubonisa ukuthi ososayensi mbumbulu bayakwenza lokhu kakade.
Izithombe ezenziwe nge-AI ezibonwe emaphepheni?
Ngo-2019, i-DARPA yethula i-Semantic Forensics yayo (SemaFor) uhlelo, ukuxhasa abacwaningi ngezimali abakha amathuluzi e-forensic akwazi ukubona imidiya eyenziwe nge-AI, ukuze kuliwe ne-disinformation.
Umkhulumeli wenhlangano ecwaninga ngezokuvikela kaMalume Sam ukuqinisekisile ukuthi ibone izithombe zezokwelapha ezingamanga ezishicilelwe emaphepheni esayensi angempela abonakala ekhiqizwa kusetshenziswa i-AI. Ngaphambi kwamamodeli ombhalo uye esithombeni, amanethiwekhi akhiqizayo aphikisanayo ayedumile. I-DARPA ibone ukuthi lawa mamodeli, aziwa kakhulu ngekhono lawo lokudala ama-deepfakes, angakha futhi izithombe zokuskena kwezokwelapha, amaseli, noma ezinye izinhlobo zezithombe ezivame ukutholakala ezifundweni ze-biomedical.
“Isimo esisongelayo sihamba ngokushesha,” kusho uWilliam Corvey, umphathi wezinhlelo zeSemaFor Irejista. “Ubuchwepheshe buya butholakala yonke indawo ngezinhloso ezinhle.” U-Corvey uthe i-ejensi ibe nempumelelo ekuthuthukiseni isofthiwe ekwazi ukuthola izithombe ezenziwe nge-GAN, futhi amathuluzi asathuthukiswa.
Indawo esongelayo ihamba ngokushesha okukhulu
“Sinemiphumela ephakamisa ukuthi ungathola ‘izingane zakini noma abazala abakude’ bendlela yokukhiqiza ofunde ukuyithola ngaphambilini, kungakhathaliseki okuqukethwe kwezithombe ezikhiqiziwe. I-SemaFor analytics ibheka izici ezihlukahlukene nemininingwane ehlotshaniswa nemidiya ekhohlisayo. , yonke into kusukela kumethadatha, okudidayo kwezibalo, kuya ekubonisweni okwengeziwe,” esho.
Abanye abahlaziyi bezithombe abahlolisisa idatha emaphepheni esayensi nabo bathole okubukeka njengezithombe ezikhiqizwe i-GAN. I-GAN iyinethiwekhi ekhiqizayo yokuphikisana, uhlobo lwesistimu yokufunda ngomshini engakwazi ukukhiqiza ukubhala, umculo, izithombe, nokuningi.
Isibonelo, uJennifer Byrne, uprofesa we-oncology yamangqamuzana eNyuvesi yaseSydney, kanye noJana Christopher, umhlaziyi wobuqotho bezithombe womshicileli wephephabhuku i-EMBO Press, bathole isethi eyinqaba yezithombe ezivela ezifundweni eziyi-17 ezihlobene ne-bio-chemistry.
Izithombe zibonise uchungechunge lwamabhendi aziwa kakhulu ngokuthi amabala asentshonalanga, okubonisa ukuba khona kwamaprotheni athile kusampula, wonke abonakala ngokumangalisayo anesizinda esifanayo. Lokho akufanele kwenzeke.
Izibonelo zengemuva eliphindayo ezithombeni ze-Western blot, ezigqanyiswe izinhlaka ezibomvu neziluhlaza … Umthombo: Byrne, uChristopher 2020
Ngo-2020, u-Byrne noChristopher bafinyelela esiphethweni sokuthi izithombe ezisolisayo cishe zakhiqizwa njengengxenye yomshini wokugaya amaphepha: umzamo wokukhiqiza amaphepha ngobuningi ocwaningweni lwamakhemikhali e-bio-chemical kusetshenziswa idatha engumgunyathi, futhi zenze ukuthi zibuyekezwe ontanga futhi zishicilelwe. I-caper enjalo ingase ihoxiswe, ngokwesibonelo, ukuze kuzuze izifundiswa ezinxeshezelwa ngokusekelwe ekuphumeni kwazo kwephepha elamukelwe, noma ukusiza umnyango ukuthi ufinyelele isilinganiso semibiko eshicilelwe.
“Ama-blots esibonelweni esiboniswe ku iphepha lethu kungenzeka ukuthi enziwe ngekhompyutha,” kusho uChristopher Irejista.
Ngivame ukuhlangana nezithombe ezibukeka zingelona iqiniso, ikakhulukazi amabala asentshonalanga, kodwa ziya ngokuya ziba nezithombe ze-microscopy
“Amaphepha okuhlola ngaphambi nangemuva kokushicilelwa, ngivame ukuhlangana nezithombe ezibukeka zingelona iqiniso, ikakhulukazi ama-blots asentshonalanga, kodwa ngokwandayo nezithombe ze-micrscopy. Ngazi kahle ukuthi eziningi zazo cishe zikhiqizwa kusetshenziswa ama-GAN.”
U-Elisabeth Bik, i-sleuth yesithombe esizimele, uyakwazi ukubona lapho izithombe zishintshiwe, naye. Ubheka imibhalo yesandla yephepha yesayensi, ezingela izithombe eziyimpinda, futhi ahlabe umkhosi lezi zindaba ukuze abahleli bamaphephabhuku bazihlole kabanzi. Kodwa kunzima ukulwa nezithombe ezingamanga uma zenziwe ngokuphelele i-algorithm.
Uveze ukuthi nakuba ingemuva eliphindaphindiwe ezithombeni eziqokonyiswe ocwaningweni luka-Byrne no-Christopher kuwuphawu lokukhohlisa, ama-blots angempela asentshonalanga ngokwawo ahlukile. Isofthiwe yombono wekhompuyutha u-Bik ayisebenzisayo ukuskena amaphepha nokubona ukukhwabanisa kwesithombe kungakuthola kunzima ukuhlaba umkhosi lawa mabhendi ngenxa yokuthi azikho izimpinda zama-blots wangempela.
“Ngeke sikuthole ukugqagqana. Wonke, ngiyakholwa, enziwe ngokwenziwa. Kanjani kahle, anginasiqiniseko,” esho etshela Irejista.
Kulula ukwenza izithombe zomgunyathi ngamamodeli akamuva e-AI akhiqizayo
Ama-GAN asuswe kakhulu amamodeli wokusabalalisa. Lawa masistimu akhiqiza izithombe ezihlukile futhi anika amandla isofthiwe yanamuhla yokuguqula umbhalo ube yisithombe ehlanganisa i-DALL-E, i-Stable Diffusion, ne-Midjourney. Bafunda ukwenza imephu yokumelwa okubonakalayo kwezinto nemiqondo baye olimini lwemvelo, futhi bangehlisa kakhulu umgoqo wokukopela kwezemfundo.
Ososayensi bangachaza nje ukuthi yiluphi uhlobo lwedatha engamanga abafuna ukwenziwa, futhi la mathuluzi azobenzela kona. Okwamanje, nokho, abakwazi ukudala izithombe zesayensi ezibukeka ngokwangempela okwamanje. Kwesinye isikhathi amathuluzi akhiqiza amaqoqo amaseli abukeka ekholisa ekuqaleni, kodwa ahluleke ngokudabukisayo uma kukhulunywa ngamabala asentshonalanga.
Lolu uhlobo lwento lezi zinhlelo ze-AI ezingakhiqiza:
Nakhu @OpenAII-DALL-E ka-DALL-E yenza ngemiyalo yeseli yebhayoloji
Ngokuqondile: “amaseli angaphansi kwesibonakhulu” kanye “namaseli T angaphansi kwesibonakhulu se-electron yokuskena” pic.twitter.com/BgcZr3k5Q5
— Tara Basu Trivedi (@tbt94) Agasti 23, 2022
UWilliam Gibson – udokotela-usosayensi kanye nozakwethu we-oncology yezokwelapha, hhayi umbhali odumile – unezinye izibonelo laphaokuhlanganisa ukuthi amamodeli anamuhla alwa kanjani nomqondo we-Western blot.
Ubuchwepheshe buya ngokuba ngcono, nokho, njengoba abathuthukisi beqeqesha amamodeli amakhudlwana ngedatha eyengeziwe.
UDavid Bimler, omunye uchwepheshe wokuqaphela ukukhohlakala kwezithombe emaphepheni esayensi, owaziwa kangcono ngokuthi uSmut Clyde, usitshele: “Abenzi bamaphepha bazofanekisa imikhiqizo yabo besebenzisa noma iyiphi indlela eshibhile futhi eshesha kakhulu, bethembele ebuthakathakeni benqubo yokubuyekezwa kontanga.”
“Bangamane bakopishe [western blots] kusuka emaphepheni amadala kodwa nalokho kuhlanganisa umsebenzi wokusesha amaphepha amadala. Okwamanje, ngiyasola, ukusebenzisa i-GAN kusengumzamo othile. Yize lokho kuzoshintsha,” engeza.
I-DARPA manje isibheke ukukhulisa uhlelo lwayo lwe-SemaFor ukuze ifunde izinhlelo zombhalo-kuya-isithombe. “Lezi zinhlobo zamamodeli zintsha impela futhi ngenkathi zisasebenza, aziyona ingxenye yomsebenzi wethu wamanje kuSemaFor,” kusho uCorvey.
“Kodwa-ke, abahloli be-SemaFor kungenzeka babheke lawa mamodeli ngesikhathi sesigaba esilandelayo sokuhlola sohlelo oluqala ngo-Fall 2023.”
Ngaleso sikhathi, ikhwalithi yocwaningo lwesayensi izoguguleka uma abashicileli bezemfundo bengakwazi ukuthola izindlela zokuthola izithombe mbumbulu ezikhiqizwe yi-AI emaphepheni. Esimeni esihle kakhulu, lolu hlobo lokukhwabanisa kwezemfundo luzokhawulelwa ezinhlelweni zokugaya amaphepha ezinganakwa kakhulu. Esimeni esibi kakhulu, kuzoba nomthelela ngisho namajenali anedumela elihle futhi ososayensi abanezinhloso ezinhle bazomosha isikhathi nemali bejaha imibono engamanga abakholelwa ukuthi iyiqiniso. ®
