Isiqulatho[Fihla][Bonisa]
I-Artificial Intelligence (AI) ekuqaleni yayicingelwa ukuba liphupha elikude, iteknoloji yexesha elizayo, kodwa oko akusekho.
Yintoni eyayisakuba sisihloko sophando ngoku iqhuma kwihlabathi lenene. I-AI ngoku ifumaneka kwiindawo ngeendawo, kubandakanya indawo osebenza kuyo, isikolo, iibhanki, izibhedlele, kunye nefowuni yakho.
Zingamehlo ezithuthi eziziqhubayo, amazwi kaSiri kunye ne-Alexa, iingqondo emva kokubikezela kwemozulu, izandla emva kokuhlinzwa kwerobhothi, kunye nokunye.
basemoyeni (AI) isiba yinto eqhelekileyo kubomi banamhlanje. Kwiminyaka eliqela edlulileyo, i-AI iye yavela njengomdlali ophambili kuluhlu olubanzi lwetekhnoloji ye-IT.
Okokugqibela, inethiwekhi ye-neural isetyenziswa yi-AI ukufunda izinto ezintsha.
Ke namhlanje siza kufunda ngeNeural Networks, indlela esebenza ngayo, iindidi zazo, usetyenziso, nokunye okuninzi.
Yintoni iNeural Network?
In yokufunda umatshini, inethiwekhi ye-neural yi-software-programmed network of artificial neurons. Izama ukuxelisa ingqondo yomntu ngokuba neentlobo ezininzi zeeseli “zemithambo-luvo,” ezifana nemithambo-luvo ebuchotsheni bethu.
Uluhlu lokuqala lwe-neurons luya kwamkela iifoto, ividiyo, isandi, isicatshulwa, kunye namanye amagalelo. Le datha ihamba kuwo onke amanqanaba, kunye nemveliso yomaleko omnye ukuya kwelinye. Oku kubaluleke kakhulu kweyona misebenzi inzima, njengokusetyenzwa kolwimi lwendalo yokufunda koomatshini.
Nangona kunjalo, kwezinye iimeko, ukujolisa ukunyanzeliswa kwenkqubo ukunciphisa ubungakanani bemodeli ngelixa kugcinwa ukuchaneka kunye nokusebenza kakuhle kuyakhethwa. Ukuthena inethiwekhi ye-neural yindlela yoxinzelelo ebandakanya ukususa iintsimbi kwimodeli efundiweyo. Qwalasela i-artificial intelligence neural network eqeqeshelwe ukwahlula abantu kwizilwanyana.
Umfanekiso uya kwahlulwa ube ngamacandelo aqaqambileyo kwaye amnyama ngoluhlu lokuqala lwe-neurons. Le datha iya kudluliselwa kuluhlu olulandelayo, oluya kugqiba apho imida.
Umaleko olandelayo uya kuzama ukuqaphela iifom eziveliswe yimiphetho. Ngokwedatha ebiqeqeshwe kuyo, idatha iya kudlula kwiileya ezininzi ngendlela efanayo ukumisela ukuba umfanekiso owubonisileyo ngowomntu okanye isilwanyana.
Xa idatha inikwe kwinethiwekhi ye-neural, iqala ukuyiqhuba. Emva koko, idatha icutshungulwa ngamanqanaba ayo ukufumana isiphumo esifunekayo. Inethiwekhi ye-neural ngumatshini ofunda kwigalelo elicwangcisiweyo kwaye ubonise iziphumo. Zintathu iindidi zokufunda ezinokuthi zenzeke kuthungelwano lwe-neural:
- Isifundo esiLawulwayo - Amagalelo kunye neziphumo zinikezelwa kwii-algorithms kusetyenziswa idatha ebhaliweyo. Emva kokuba befundiswe indlela yokuhlalutya idatha, baqikelela isiphumo ekujoliswe kuso.
- Ukufunda okungajongwanga-I-ANN ifunda ngaphandle koncedo lomntu. Akukho datha ibhaliweyo, kwaye imveliso igqitywe ngeepateni ezifunyenwe kwidatha yemveliso.
- Ukomeleza ukuFunda kuxa inethiwekhi ifunda kwingxelo eyifumanayo.
Zisebenza njani iNeural network?
Iineuron ezenziweyo zisetyenziswa kwiinethiwekhi ze-neural, eziziinkqubo ezintsonkothileyo. I-neurons eyenziweyo, eyaziwa ngokuba yi-perceptrons, yenziwe ngala malungu alandelayo:
- input
- ubunzima
- Bias
- Umsebenzi wokuqalisa
- imveliso
Iimaleko zemithambo-luvo ezenza uthungelwano lwe-neural. Inethiwekhi ye-neural inemigangatho emithathu:
- Umaleko wegalelo
- Umaleko ofihlakeleyo
- Imveliso umaleko
Idatha kwifom yexabiso lamanani ithunyelwa kuluhlu lwegalelo. Iileya ezifihliweyo zenethiwekhi zizona zenza izibalo ezininzi. Umaleko wemveliso, okokugqibela kodwa okungancinci, uqikelela isiphumo. IiNeurons zilawula enye kwenye kwinethiwekhi ye-neural. IiNeurons zisetyenziselwa ukwakha umaleko ngamnye. Idatha ihanjiswa kumaleko afihliweyo emva kokuba igalelo lengeniso liyifumene.
Ubunzima busetyenziswa kwigalelo ngalinye. Ngaphakathi kweeleya ezifihliweyo zenethiwekhi ye-neural, ubunzima lixabiso eliguqulela idatha engenayo. Ubunzima busebenza ngokuphinda-phinda idatha yegalelo ngexabiso lobunzima kumaleko wokufaka.
Iqala ke ixabiso lokuqala elifihliweyo. Idatha yegalelo iyaguqulwa kwaye idluliselwe komnye umaleko ngokusebenzisa iileya ezifihliweyo. Uluhlu lwemveliso luxanduva lokuvelisa umphumo wokugqibela. Amagalelo kunye neentsimbi ziyaphindaphindwa, kwaye umphumo unikezelwa kwi-neuron efihliweyo njengesamba. I-neuron nganye inikwe icala. Ukubala itotali, i-neuron nganye yongeza amagalelo ewafumanayo.
Emva koko, ixabiso lidlula ngomsebenzi wokuvula. Isiphumo somsebenzi wokuvula sinquma ukuba i-neuron iyasebenza okanye hayi. Xa i-neuron isebenza, ithumela ulwazi kwezinye iileya. Idatha yenziwe kuthungelwano de i-neuron ifikelele kwinqanaba lemveliso isebenzisa le ndlela. Ukusasaza phambili lelinye igama lale nto.
Indlela yokondla idatha kwi-node yegalelo kunye nokufumana imveliso nge-node yemveliso eyaziwa ngokuba yi-feed-forward propagation. Xa idatha yegalelo yamkelwa ngumaleko ofihliweyo, ukusasazwa kwe-feed-forward kwenzeka. Icutshungulwa ngokomsebenzi wokuvula kwaye emva koko idluliselwe kwimveliso.
Isiphumo siqikelelwa yi-neuron kumaleko wemveliso ngeyona ndlela iphezulu yokunokwenzeka. Ukusasazwa ngasemva kwenzeka xa imveliso ingachanekanga. Ubunzima buqaliswa kwigalelo ngalinye ngelixa kusenziwa inethiwekhi ye-neural. I-backpropagation yinkqubo yokulungiswa kwakhona kobunzima begalelo ngalinye ukunciphisa iimpazamo kunye nokubonelela ngesiphumo esichanekileyo.
Iindidi zeNeural Network
1. Perceptron
Imodeli ye-Minsky-Papert perceptron yenye yezona modeli zilula kwaye zindala ze-neuron. Yeyona yunithi incinci yenethiwekhi ye-neural eyenza izibalo ezithile ukuze kufunyanwe iimpawu okanye ubukrelekrele beshishini kwidatha engenayo. Kuthatha amagalelo anobunzima kwaye isebenzise umsebenzi wokuvula ukufumana umphumo wokugqibela. TLU (threshold logic unit) lelinye igama leperceptron.
I-Perceptron yi-binary classifier eyinkqubo yokufunda egadiweyo eyahlula idatha ibe ngamaqela amabini. Amasango eLogic ezifana AND, OKANYE, kunye NAND inokuphunyezwa nge perceptrons.
2. I-Feed-Forward Neural Network
Olona guqulelo lusisiseko lothungelwano lwe-neural, apho idatha yegalelo ihamba ngokukodwa kwicala elinye, idlula kwii-neural nodes ezenziweyo kwaye iphuma ngeenodi zemveliso. Iileya zegalelo kunye nemveliso zikhona kwiindawo apho iileya ezifihlakeleyo zinokuthi zibekho okanye zingabikho. Zinokuphawulwa njengonaleko olunye okanye uthungelwano olunolwaleko oluninzi lwe-neural olusekwe koku.
Inani leeleya ezisetyenzisiweyo lichongwa kukuntsonkotha komsebenzi. Isasaza kuphela phambili kwicala elinye kwaye ayisasazeki ngasemva. Apha, iintsimbi zihlala zingatshintshi. Amagalelo aphindaphindwa ngobunzima ukondla umsebenzi wokuvula. Umsebenzi wokwenza kusebenze okanye inyathelo lokuvula inyathelo liyasetyenziswa ukwenza oku.
3. I-perceptron yamanqanaba amaninzi
Intshayelelo ephucukileyo iminatha ye-neural, apho idatha yegalelo ihanjiswa ngeendlela ezininzi zee-neurons ezenziweyo. Yinethiwekhi ye-neural eqhagamshelwe ngokupheleleyo, kuba yonke i-node iqhagamshelwe kuzo zonke ii-neuron kolu luhlu lulandelayo. Iileya ezininzi ezifihliweyo, oko kukuthi, ubuncinci iileya ezintathu okanye ngaphezulu, zikhona kwigalelo kunye nemveliso yomaleko.
Inokusasazwa kwe-bidirectional, okuthetha ukuba inokusasaza phambili nangasemva. Amagalelo aphindaphindwa ngobunzima kwaye athunyelwe kumsebenzi wokuvula, apho atshintshwa ngokusasazwa ngasemva ukuze kuncitshiswe ilahleko.
Ubunzima ngamaxabiso afundiwe ngoomatshini kwiNeural Networks, ukuyibeka ngokulula. Ngokuxhomekeke kumahluko phakathi kweziphumo ezilindelekileyo kunye namagalelo oqeqesho, ziyazilungisa ngokwazo. ISoftmax isetyenziswa njengomsebenzi wokwenza umaleko wokuphuma emva kwemisebenzi engasebenziyo.
4. Convolutional Neural Network
Ngokuchaseneyo noluhlu oluqhelekileyo lwe-dimensional, inethiwekhi ye-convolution ye-neural ine-demensional configuration ye-neurons. Umaleko wokuqala waziwa ngokuba yi-convolutional layer. I-neuron nganye ekumaleko we-convolution yenza kuphela ulwazi olusuka kwinxalenye elinganiselweyo yendawo yokubonwayo. Njengesihluzi, iimpawu zongeniso zithathwa kwimowudi yebhetshi.
Inethiwekhi iyayiqonda imifanekiso kumacandelo kwaye inokwenza ezi ntshukumo amaxesha amaninzi ukugqiba ukusetyenzwa komfanekiso.
Umfanekiso uguqulwa ukusuka kwi-RGB okanye kwi-HSI ukuya kwi-greyscale ngexesha lokucubungula. Ukwahluka okungaphezulu kwixabiso lepixel kuya kunceda ekuchongeni imiphetho, kwaye imifanekiso ingahlelwa ngokwamaqela aliqela. Ukusasazwa kwe-unidirectional kwenzeka xa i-CNN iqulethe i-convolutional layers enye okanye ngaphezulu elandelwa yi-pooling, kunye ne-bidirectional propagation yenzeke xa ukukhutshwa kwe-convolution layer ithunyelwa kwi-neural network edityaniswe ngokupheleleyo yokuhlela umfanekiso.
Ukukhupha izinto ezithile zomfanekiso, izihluzi ziyasetyenziswa. Kwi-MLP, amagalelo anobunzima kwaye anikezelwe kumsebenzi wokuvula. I-RELU isetyenziswe kwi-convolution, ngelixa i-MLP isebenzisa umsebenzi ongasebenziyo ongasebenziyo olandelwa yi-softmax. Kumfanekiso kunye nokuqatshelwa kwevidiyo, ulwahlulo lwesemantic, kunye nokufunyanwa kweparaphrase, uthungelwano lwe-neural convolutional luvelisa iziphumo ezigqwesileyo.
5. Inethiwekhi yeRadial Bias
I-vector ye-input ilandelwa ngumaleko we-RBF neurons kunye ne-output layer kunye ne-node enye kwicandelo ngalinye kwi-Radial Basis Function Network. Igalelo lihlelwa ngokuthelekisa ngokuchasene namanqaku edatha ukusuka kwiseti yoqeqesho, apho i-neuron nganye igcina iprototype. Lo ngomnye wemizekelo yeeseti zoqeqesho.
I-neuron nganye ibala umgama we-Euclidean phakathi kwegalelo kunye neprototype yayo xa ivektha entsha yegalelo [ivektha engu-n-dimensional ozama ukuyihlela] kufuneka ihlelwe. Ukuba sineeklasi ezimbini, iKlasi A kunye neKlasi B, igalelo elitsha eliza kuhlulwa lifane kakhulu neprototypes yodidi A kuneprototypes yodidi B.
Ngenxa yoko, inokuphawulwa okanye ihlelwe njengodidi A.
6. INeural Network yeRecurrent
INeural Neural Networks ziyilelwe ukuba zigcine imveliso yomaleko kwaye emva koko uyondle ukubuyisela kwigalelo ukunceda uqikelelo lweziphumo zomaleko. Umphakeli-phambili inethiwekhi yomnatha iqhele ukuba ngumaleko wokuqala, olandelwa ngumaleko wothungelwano lwe-neural oluqhelekileyo, apho umsebenzi wenkumbulo ukhumbula inxalenye yolwazi ebinalo kwinyathelo lexesha elidlulileyo.
Lo mzekelo usebenzisa ukusasazwa kwangaphambili. Igcina idatha eya kufuneka kwixesha elizayo. Kwimeko apho ingqikelelo ayichanekanga, ireyithi yokufunda isetyenziselwa ukwenza uhlengahlengiso olungephi. Ngenxa yoko, njengoko i-backpropagation iqhubeka, iya kuba ichaneka ngakumbi.
izicelo
Iinethiwekhi ze-Neural zisetyenziselwa ukujongana neengxaki zedatha kwiindidi ezahlukeneyo zezifundo; eminye imizekelo iboniswe ngezantsi.
- Ukuqwalaselwa koBuso - Izisombululo zoKubonwa koBuso zisebenza njengeenkqubo zokucupha ezisebenzayo. Iinkqubo zokuqaphela zinxulumanisa iifoto zedijithali kubuso babantu. Zisetyenziswa kwiiofisi ekungeneni okukhethiweyo. Ngaloo ndlela, iinkqubo ziqinisekisa ubuso bomntu kwaye zithelekise noluhlu lwee-ID ezigcinwe kwisiseko sayo sedatha.
- I-Stock Prediction - Utyalo-mali luvezwe kwimingcipheko yemarike. Kunzima ukubona kwangaphambili uphuhliso lwexesha elizayo kwimarike yemasheya eguquguqukayo kakhulu. Ngaphambi kothungelwano lwe-neural, izigaba ze-bullish kunye ne-bearish ezihlala ziguquka zazingalindelekanga. Kodwa, yintoni eyatshintsha yonke into? Ewe, sithetha ngothungelwano lwe-neural… I-Multilayer Perceptron MLP (uhlobo lwenkqubo yobuntlola eyenziweyo ethunyelwa phambili) isetyenziselwa ukwenza uqikelelo lwesitokhwe esiyimpumelelo ngexesha lokwenyani.
- Imidiya yokuncokola -Nokuba ivakala kangakanani na, imidiya yoluntu itshintshile indlela yokuphila. Ukuziphatha kwabasebenzisi beendaba zoluntu kufundwa kusetyenziswa iiNethiwekhi zeNeural Artificial. Uhlalutyo olukhuphisanayo, idatha enikezelwa yonke imihla ngonxibelelwano olubonakalayo luyafunjwa kwaye luvavanywe. Izenzo zabasebenzisi beendaba zoluntu ziphindaphindwa yi-neural networks. Iindlela zokuziphatha zomntu ngamnye zinokuqhagamshelwa kwiipatheni zenkcitho yabantu nje ukuba idatha ihlalutywe ngeintanethi zemidiya yoluntu. Idatha evela kwizicelo zemidiya yoluntu ichithwa kusetyenziswa i-Multilayer Perceptron ANN.
- Ukhathalelo lwempilo – Abantu kwihlabathi lanamhlanje basebenzisa izibonelelo zetekhnoloji kwishishini lezempilo. Kwishishini lezempilo, iiNethiwekhi ze-Convolutional Neural zisetyenziselwa ukufumanisa i-X-ray, i-CT scans, kunye ne-ultrasound. Idatha ye-imaging yezonyango efunyenwe kwiimvavanyo ezikhankanywe ngasentla ivavanywa kwaye ihlolwe kusetyenziswa imodeli yenethiwekhi ye-neural, njengoko i-CNN isetyenziselwa ukulungiswa kwemifanekiso. Kuphuhliso lweenkqubo zokuqaphela ilizwi, inethiwekhi ye-neural ephindaphindiweyo (RNN) iphinda isetyenziswe.
- Ingxelo yeMozulu – Phambi kokuphunyezwa kobukrelekrele bokwenziwa, uqikelelo lwesebe lezemozulu aluzange luchaneke. Uqikelelo lwemozulu lwenziwa ubukhulu becala ukuqikelela iimeko zemozulu eziya kwenzeka kwixesha elizayo. Uqikelelo lwemozulu lusetyenziswa ukuqikelela ukubakho kweentlekele zendalo kwixesha langoku. Uqikelelo lwemozulu lwenziwa kusetyenziswa i-multilayer perceptron (MLP), convolutional neural networks (CNN), kunye nerecurrent neural networks (RNN).
- Ukhuselo-Ulungiselelo, uhlalutyo lohlaselo oluxhobileyo, kunye nendawo yento zonke zisebenzisa uthungelwano lwe-neural. Bakwaqeshwa kwiipatroli zasemoyeni naselwandle, kunye nokulawula iidrones ezizimeleyo. Ubukrelekrele bokwenziwa bunika ishishini lezokhuselo ukomelela olufunekayo ukuze linyuse itekhnoloji yalo. Ukubona ubukho bemigodi engaphantsi kwamanzi, iConvolutional Neural Networks (CNN) iyasetyenziswa.
eziluncedo
- Nokuba ii-neuron ezimbalwa kuthungelwano lwe-neural azisebenzi kakuhle, uthungelwano lwe-neural lusezakuvelisa iziphumo.
- Iinethiwekhi ze-Neural ziyakwazi ukufunda ngexesha lokwenyani kwaye zilungelelanise nokutshintsha kwazo iisetingi.
- Iinethiwekhi zeNeural zinokufunda ukwenza imisebenzi eyahlukeneyo. Ukubonelela ngesiphumo esichanekileyo ngokusekelwe kwidatha enikiweyo.
- Uthungelwano lweNeural lunamandla kunye nokukwazi ukuphatha imisebenzi emininzi ngexesha elinye.
nezingeloncedo
- Iinethiwekhi zeNeural zisetyenziselwa ukusombulula iingxaki. Ayichazi inkcazo emva kokuba "kutheni kwaye njani" yenza izigwebo ezenzileyo ngenxa yokuntsonkotha kothungelwano. Ngenxa yoko, ukuthembana kwenethiwekhi kunokupheliswa.
- Amalungu enethiwekhi ye-neural axhomekeke kwelinye. Oko kukuthi, iinethiwekhi ze-neural zifuna (okanye zixhomekeke kakhulu) iikhompyuter ezinamandla aneleyo ekhompyuter.
- Inkqubo yenethiwekhi ye-neural ayinamgaqo othile (okanye umthetho wesithupha). Kwindlela yovavanyo kunye nempazamo, ulwakhiwo lwenethiwekhi oluchanekileyo lusekiwe ngokuzama eyona nethiwekhi ilungileyo. Yinkqubo efuna ukulungiswa kakuhle kakhulu.
isiphelo
Intsimi amanethiwekhi isanda ngokukhawuleza. Kubalulekile ukufunda nokuqonda iikhonsepthi kweli candelo ukuze ukwazi ukujongana nazo.
Iindidi ezininzi zothungelwano lwe-neural ziye zagutyungelwa kweli nqaku. Unokusebenzisa iinethiwekhi ze-neural ukujongana neengxaki zedatha kweminye imimandla ukuba ufunda ngakumbi ngolu qeqesho.
Shiya iMpendulo