I-AI inamandla okuphucula ukusebenza kakuhle kumacandelo ahlukeneyo afana neshishini kunye nokhathalelo lwempilo. Nangona kunjalo, ukunqongophala kokuchazwa kuthintela ukuthembela kwethu ekuyisebenziseni ukwenza izigqibo.
Ngaba sifanele sithembele kwisigwebo se-algorithm?
Kubalulekile ukuba abenzi bezigqibo kulo naliphi na ishishini baqonde imida kunye nokuthambekela okunokwenzeka iimodeli zokufunda ngomatshini. Ukuqinisekisa ukuba le mizekelo iziphatha ngendlela ecetywayo, ukuveliswa kwayo nayiphi na inkqubo ye-AI kufuneka icaciswe kumntu.
Kweli nqaku, siya kuhamba ngokubaluleka kokuchazwa kwe-AI. Siza kubonelela ngokufutshane kwiindidi zeendlela ezisetyenziselwa ukufumana iinkcazo kwiimodeli zokufunda koomatshini.
Yintoni i-AI echazwayo?
Iyachazwa kukubhadla okungeyonyani okanye i-XAI ibhekisa kubuchule kunye neendlela ezisetyenziswayo ukuvumela abantu ukuba baqonde ukuba imifuziselo yokufunda koomatshini ifika njani kwisiphumo esithile.
Zininzi ezaziwayo umatshini wokufunda iialgorithms sebenza ngokungathi "yibhokisi emnyama". Ngokufunda koomatshini, ibhokisi emnyama algorithms bhekisa kwiimodeli zeML apho kungenzekiyo ukuqinisekisa ukuba igalelo elithile likhokelela njani kwimveliso ethile. Nangona umphuhlisi we-AI akayi kukwazi ukuchaza ngokupheleleyo indlela i-algorithm isebenza ngayo.
Ngokomzekelo, ukusetyenziswa kwe-algorithms yokufunda nzulu amanethiwekhi ukuchonga iipateni ukusuka kwitoni yedatha. Nangona abaphandi be-AI kunye nabaphuhlisi beyiqonda indlela uthungelwano lwe-neural olusebenza ngayo ngokwembono yezobugcisa, bengenako nokuchaza ngokupheleleyo ukuba inethiwekhi ye-neural ize njani nesiphumo esithile.
Ezinye iinethiwekhi ze-neural ziphatha izigidi zeeparamitha ezithi zonke zisebenze kunye ukubuyisela iziphumo zokugqibela.
Kwiimeko apho izigqibo zibalulekile, ukunqongophala kokuchazwa kunokuba yingxaki.
Kutheni Kubalulekile Ukucacisa
Ukuchazwa inika ingqiqo kwindlela iimodeli zenza izigqibo. Amashishini aceba ukulungelelanisa i-AI ukuze enze izigqibo kuya kufuneka anqume ukuba i-AI isebenzise igalelo elifanelekileyo ukuze ifike kwesona sigqibo silungileyo.
Iimodeli ezingenakuchazwa ngumba kumashishini amaninzi. Umzekelo, ukuba inkampani ibinokusebenzisa ialgorithm ukwenza izigqibo zokuqesha, bekuya kuba yeyona nto ilungileyo kumntu wonke ukuba nokufihlwa kwindlela ialgorithm egqiba ngayo ukwala umfaki-sicelo.
Enye indawo apho ukufunda okunzulu Ii-algorithms zisetyenziswa rhoqo kukhathalelo lwempilo. Kwiimeko apho i-algorithms izama ukufumanisa iimpawu ezinokuthi zibe nomhlaza, kubalulekile ukuba oogqirha baqonde indlela imodeli efike ngayo kuxilongo oluthile. Inqanaba elithile lokucaciswa liyafuneka ukuba iingcali zisebenzise ngokupheleleyo i-AI kwaye zingayilandeli ngokumfamekileyo
Isishwankathelo se-AI Algorithms echazwayo
Ii-algorithms ze-AI ezichazwayo ziwela kwiindidi ezimbini ezibanzi: iimodeli ezizichazayo kunye neenkcazo ze-post-hoc.
Iimodeli ezizitolikayo
Iimodeli ezizitolikayo zii-algorithms ezithi umntu azifunde ngokuthe ngqo aze azitolike. Kule meko, imodeli ngokwayo ingcaciso.
Ezinye zeemodeli eziqhelekileyo ezizitolikayo ziquka imithi yesigqibo kunye neemodeli zokubuyisela.
Umzekelo, makhe sithathele ingqalelo imodeli yohlengahlengiso yomgca eqikelela amaxabiso endlu. Ukuhlehla ngomgca kuthetha ukuba ngexabiso elithile x, siya kuba nakho ukuqikelela ixabiso lethu ekujoliswe kulo y ngokusebenzisa umsebenzi othile womda f.
Masithi imodeli yethu isebenzisa ubungakanani obukhulu njengelona galelo liphambili lokumisela ixabiso lendlu. Sisebenzisa ubuyiselo lomgca, siye sakwazi ukuza nomsebenzi y = 5000 * x apho u-x lixabiso leenyawo ezikwere okanye ubungakanani beqashiso.
Le modeli iyafundeka ngabantu kwaye iselubala ngokupheleleyo.
Iinkcazo zePost-Hoc
Iinkcazo zasemva kwe-hoc liqela lee-algorithms kunye nobuchule obunokusetyenziswa ukongeza ukucaciswa kwezinye ii-algorithms.
Uninzi lweendlela zenkcazo ye-post-hoc ayifuni ukuqonda ukuba i-algorithm isebenza njani. Umsebenzisi ufuna kuphela ukucacisa igalelo kunye nesiphumo se-algorithm ekujoliswe kuyo.
Ezi ngcaciso zohlulwe kwakhona zibe ziindidi ezimbini: iingcaciso zendawo kunye neengcaciso zehlabathi.
Iingcaciso zasekuhlaleni zijolise ekuchazeni isethi engaphantsi yamagalelo. Umzekelo, xa kunikwe isiphumo esithile, ingcaciso yendawo iya kuba nakho ukukhomba ukuba yeyiphi iparameters ebenegalelo ekwenzeni eso sigqibo.
Iinkcazo zehlabathi jikelele zijolise ekuveliseni iinkcazo ze-post-hoc ze-algorithm yonke. Olu hlobo lwenkcazo lukholisa ukuba nzima ukwenza. Ii-algorithms zintsonkothile kwaye kusenokubakho iiparitha ezingenakubalwa ezibalulekileyo ekuphumezeni isiphumo sokugqibela.
Imizekelo yeeNgcaciso zeNdawo zeNdawo
Phakathi kweendlela ezininzi ezisetyenzisiweyo ukufezekisa i-XAI, i-algorithms esetyenziselwa iinkcazo zengingqi yinto uninzi lwabaphandi bagxile kuyo.
Kweli candelo, siza kujonga ezinye iialgorithms ezidumileyo zenkcazo yendawo kunye nendlela nganye yazo esebenza ngayo.
IXESHA
I-LIME (Imodeli yasekuhlaleni eTolikayo-i-Agnostic Explainer) yi-algorithm enokuthi ichaze ukuqikelelwa kwayo nayiphi na i-algorithm yokufunda umatshini.
Njengoko igama lisitsho, i-LIME yimodeli-agnostic. Oku kuthetha ukuba i-LIME inokusebenzela naluphi na uhlobo lwemodeli. Imodeli ikwanokutolika ekuhlaleni, okuthetha ukuba sinokuyichaza imodeli sisebenzisa iziphumo zasekhaya endaweni yokuchaza yonke imodeli.
Nokuba imodeli echazwayo yibhokisi emnyama, i-LIME idala imodeli yomgca wendawo malunga neengongoma ezikufutshane nendawo ethile.
I-LIMe ibonelela ngemodeli yomgca esondele kwimodeli kwindawo yoqikelelo kodwa hayi kwihlabathi liphela.
Unokufunda ngakumbi malunga nale algorithm ngokundwendwela lo vimba womthombo ovulekileyo.
SHAP
I-Shapley Additive Explanations (SHAP) yindlela yokuchaza uqikelelo lomntu ngamnye. Ukuqonda indlela esebenza ngayo i-SHAP, kuya kufuneka sichaze ukuba zeziphi iinqobo zeShapley.
Ixabiso le-Shapley yingqiqo kwithiyori yomdlalo ebandakanya ukunika "ixabiso" kumdlali ngamnye kumdlalo. Oku kwabiwa ngendlela yokuba ixabiso elabelwe umdlali ngamnye lisekelwe kwigalelo lomdlali kumdlalo.
Sifaka njani isicelo ithiyori yomdlalo kumatshini wokufunda iimodeli?
Masithi into nganye kwimodeli yethu "ngumdlali" kwaye "umdlalo" ngumsebenzi ovelisa uqikelelo.
Indlela ye-SHAP yenza imodeli yomgca enobunzima enika ixabiso le-Shapley kwiimpawu ezahlukeneyo. Iimpawu ezinamaxabiso aphezulu eShapley zinempembelelo enkulu kwisiphumo semodeli ngelixa iimpawu ezinamaxabiso aphantsi eShapley zinempembelelo encinci.
isiphelo
Ukucaciswa kwe-AI akubalulekanga kuphela ekuqinisekiseni ubulungisa kunye nokuphendula kweenkqubo ze-AI, kodwa kunye nokwakha ukuthembela kwi-teknoloji ye-AI ngokubanzi.
Kusekho uphando oluninzi oluza kwenziwa kwindawo yokuchazwa kwe-AI, kodwa kukho iindlela ezithembisayo ezinokusinceda siqonde iisistim ze-AI zebhokisi ezimnyama esele zisetyenziswa ngokubanzi namhlanje.
Ngophando oluthe kratya kunye nophuhliso, sinethemba lokwakha iinkqubo ze-AI ezicace ngakumbi kwaye kulula ukuziqonda. Okwangoku, amashishini kunye neengcali kwiinkalo ezifana nokhathalelo lwezempilo kufuneka baqonde imida yokuchazwa kwe-AI.
Shiya iMpendulo