{"id":799,"date":"2023-02-07T15:09:07","date_gmt":"2023-02-07T12:09:07","guid":{"rendered":"https:\/\/legendhomework.com\/tr\/?p=799"},"modified":"2023-02-07T15:09:07","modified_gmt":"2023-02-07T12:09:07","slug":"algoritmik-cozum-multimedya-bolumu-multimedya-bolumu-odevleri-multimedya-bolumu-tez-yaptirma-multimedya-bolumu-odev-ucretleri","status":"publish","type":"post","link":"https:\/\/legendhomework.com\/tr\/algoritmik-cozum-multimedya-bolumu-multimedya-bolumu-odevleri-multimedya-bolumu-tez-yaptirma-multimedya-bolumu-odev-ucretleri\/","title":{"rendered":"Algoritmik \u00c7\u00f6z\u00fcm \u2013 Multimedya B\u00f6l\u00fcm\u00fc \u2013 Multimedya B\u00f6l\u00fcm\u00fc \u00d6devleri \u2013 Multimedya B\u00f6l\u00fcm\u00fc Tez Yapt\u0131rma \u2013Multimedya B\u00f6l\u00fcm\u00fc \u00d6dev \u00dccretleri"},"content":{"rendered":"

Boyut Azaltma<\/span><\/strong><\/h3>\n

D b\u00fcy\u00fck oldu\u011funda, y\u00fcksek boyutlu girdi uzay\u0131ndaki e\u011fitim \u00f6rnekleri \u00e7ok seyrekle\u015fir (N > D olsa bile). Seyrek n\u00fcfuslu alan, istatistiksel \u00e7\u0131kar\u0131mlar yapmak i\u00e7in iki teorik engele neden olur.<\/span><\/p>\n

\u0130lk olarak, bir sorgu \u00f6rne\u011fi verildi\u011finde, en yak\u0131n kom\u015fular\u0131, veriler d\u00fczg\u00fcn bir \u015fekilde da\u011f\u0131t\u0131ld\u0131\u011f\u0131nda e\u015fit uzakl\u0131kta olma e\u011filimindedir. (G\u00f6r\u00fcnt\u00fc veri setlerinin \u00e7o\u011fu giri\u015f uzay\u0131nda tek tip olarak da\u011f\u0131lmad\u0131\u011f\u0131ndan, bu endi\u015fe bizim durumumuz i\u00e7in ge\u00e7erli olmayabilir.) \u0130kinci olarak, en yak\u0131n kom\u015fular sorgulama noktas\u0131na “yerel” de\u011fildir. Bu ikinci matematik problemi, s\u0131n\u0131f tahminini zorla\u015ft\u0131r\u0131r.<\/span><\/p>\n

Boyut azaltmay\u0131 ger\u00e7ekle\u015ftirmek i\u00e7in temel bile\u015fen analizi (PCA) ve ba\u011f\u0131ms\u0131z bile\u015fen analizi (ICA) gibi bir\u00e7ok y\u00f6ntem \u00f6nerilmi\u015ftir.<\/span><\/p>\n

Son zamanlarda, y\u00fcksek boyutlu bir uzaya g\u00f6m\u00fcl\u00fc do\u011frusal olmayan manifoldlar\u0131 bulmak i\u00e7in birka\u00e7 boyutluluk azaltma algoritmas\u0131 \u00f6nerilmi\u015ftir. \u00d6nerilen y\u00f6ntemler aras\u0131nda, g\u00f6r\u00fcnt\u00fc k\u00fcmeleme ve g\u00f6r\u00fcnt\u00fc alma g\u00f6revlerine Isomap yerel do\u011frusal g\u00f6mme ve \u00e7ekirdek PCA uygulanm\u0131\u015ft\u0131r.<\/span><\/p>\n

Bununla birlikte, \u00e7oklu \u00f6\u011frenmenin etkili g\u00f6r\u00fcnd\u00fc\u011f\u00fc senaryolar olduk\u00e7a yapmac\u0131k g\u00f6r\u00fcnmektedir. \u00d6rne\u011fin, yayg\u0131n olarak kullan\u0131lan Swiss-roll \u00f6rne\u011fi, verilerin yo\u011fun bir \u015fekilde dolduruldu\u011fu \u00fc\u00e7 boyutlu bir yap\u0131d\u0131r.<\/span><\/p>\n

Literat\u00fcrlerde sunulan birka\u00e7 y\u00fcz ve nesne g\u00f6r\u00fcnt\u00fcs\u00fc \u00f6rne\u011fi, bir g\u00f6r\u00fcnt\u00fcden di\u011ferine pozlar\u0131n\u0131 yaln\u0131zca biraz de\u011fi\u015ftirir, b\u00f6ylece manifoldlar kolayca ke\u015ffedilebilir. Ayr\u0131ca, g\u00f6sterilen senaryolar\u0131n hi\u00e7biri g\u00fcr\u00fclt\u00fcy\u00fc \u00e7oklu \u00f6\u011frenmeyi ciddi \u015fekilde zorlayacak bir fakt\u00f6r olarak g\u00f6rmemi\u015ftir.<\/span><\/p>\n

Sorgu-kavram \u00f6\u011frenimi ba\u011flam\u0131nda bu boyut indirgeme y\u00f6ntemlerinin en b\u00fcy\u00fck dezavantaj\u0131, boyutsall\u0131\u011f\u0131 t\u00fcm veri k\u00fcmesine ve t\u00fcm kullan\u0131c\u0131lara g\u00f6re evrensel bir modda azaltmalar\u0131d\u0131r. Boyutsall\u0131\u011f\u0131 herkese uyan tek bir \u015fekilde azaltmak, bireylerin \u00f6znelli\u011fini tamamen g\u00f6z ard\u0131 eder.<\/span><\/p>\n

\u00c7al\u0131\u015fmam\u0131z, alg\u0131sal bir mesafe fonksiyonunun sadece boyutsal olarak k\u0131smi olmad\u0131\u011f\u0131n\u0131, ayn\u0131 zamanda iki g\u00f6r\u00fcnt\u00fc aras\u0131ndaki benzerli\u011fin \u00f6l\u00e7\u00fcld\u00fc\u011f\u00fc alt uzayda da dinamik oldu\u011funu g\u00f6stermektedir. Bu nedenle, uyarlanabilir boyutluluk indirgemesi ger\u00e7ekle\u015ftirmek i\u00e7in Dinamik K\u0131smi \u0130\u015flevi (DPF) kullanmay\u0131 \u00f6neriyoruz.<\/span><\/p>\n

DPF’nin neden \u00e7al\u0131\u015ft\u0131\u011f\u0131n\u0131 k\u0131saca \u00f6zetleyelim. Bili\u015fsel psikoloji toplulu\u011fundaki baz\u0131 \u00f6nemli \u00e7al\u0131\u015fmalar, bir nesne \u00e7ifti aras\u0131ndaki benzerli\u011fi \u00f6l\u00e7erken t\u00fcm \u00f6zelliklerin dikkate al\u0131nmas\u0131 gerekmedi\u011fini g\u00f6steren \u00e7ok say\u0131da kan\u0131t sa\u011flam\u0131\u015ft\u0131r.<\/span><\/p>\n

Son zamanlarda, benzerlik alg\u0131s\u0131n\u0131n, benzerli\u011fi \u00f6l\u00e7mek i\u00e7in sayg\u0131lar\u0131 (\u00f6zellikleri veya nitelikleri) belirleyen s\u00fcre\u00e7 oldu\u011funu g\u00f6sterdi. Daha kesin olarak, bir nesne \u00e7iftini \u00f6l\u00e7mek i\u00e7in bir benzerlik fonksiyonu, kar\u015f\u0131la\u015ft\u0131rma yap\u0131lmadan \u00f6nce de\u011fil, yaln\u0131zca nesneler kar\u015f\u0131la\u015ft\u0131r\u0131ld\u0131ktan sonra form\u00fcle edilir; ve kar\u015f\u0131la\u015ft\u0131rmaya sayg\u0131, bu form\u00fclasyon s\u00fcrecinde dinamik olarak etkinle\u015ftirilir.<\/span><\/p>\n

A\u00e7\u0131klamak i\u00e7in basit bir \u00f6rnek kullanal\u0131m. Diyelim ki \u0130ngiltere’ye benzeyen iki yerin ad\u0131n\u0131 s\u00f6ylememiz isteniyor. \u00c7e\u015fitli olas\u0131l\u0131klar aras\u0131nda makul bir cevap olabilir.<\/span><\/p>\n

Bununla birlikte, \u0130ngiltere’nin \u0130sko\u00e7ya’ya benzedi\u011fi y\u00f6nler, \u0130ngiltere’nin New England’a benzedi\u011fi y\u00f6nlerden farkl\u0131d\u0131r. \u0130ngiltere ve \u0130sko\u00e7ya’n\u0131n ortak \u00f6zelliklerini \u0130ngiltere ve New England’\u0131 kar\u015f\u0131la\u015ft\u0131rmak i\u00e7in kullan\u0131rsak, ikinci \u00e7ift benzer g\u00f6r\u00fcnmeyebilir ve bunun tersi de ge\u00e7erlidir.\u00a0<\/span>Bu nedenle, sabit bir a\u00e7\u0131 k\u00fcmesi kullanan bir uzakl\u0131k i\u015flevi, farkl\u0131 a\u00e7\u0131lardan benzer olan nesneleri yakalayamaz.\u00a0<\/span><\/p>\n

\n

Algoritma<\/a> cesitleri<\/span>
\nAlgoritma Nedir<\/span>
\nAlgoritma c\u00fcmle i\u00e7inde kullan\u0131m\u0131<\/span>
\nalgoritma <\/span>
\n5 tane algoritma \u00f6rne\u011fi<\/span>
\nAlgoritma Nas\u0131l yap\u0131l\u0131r<\/span>
\nGeleneksel algoritma<\/span>
\nAlgoritma Sorular\u0131<\/span><\/p>\n

\n

E\u011fitim S\u0131n\u0131flar\u0131n\u0131n Dengesizli\u011fi<\/span><\/strong><\/h3>\n

S\u0131n\u0131fland\u0131rma s\u00fcrecini engelleyen ince ama ciddi bir sorun, dengesiz e\u011fitim verilerinin neden oldu\u011fu \u00e7arp\u0131k s\u0131n\u0131f s\u0131n\u0131r\u0131d\u0131r. Bu sorunu g\u00f6stermek i\u00e7in, bir 2B dama tahtas\u0131 \u00f6rne\u011fi kullan\u0131yoruz.<\/span><\/p>\n

D\u00f6rt \u00e7eyre\u011fe b\u00f6l\u00fcnm\u00fc\u015f 200 \u0301200 karelik bir dama tahtas\u0131 g\u00f6sterir. Sol \u00fcst ve sa\u011f alt kadranlar negatif (\u00e7o\u011funluk) \u00f6rneklerle dolu, ancak sa\u011f \u00fcst ve sol alt kadranlar yaln\u0131zca pozitif (az\u0131nl\u0131k) \u00f6rnekleri i\u00e7eriyor.<\/span><\/p>\n

S\u0131n\u0131flar aras\u0131ndaki \u00e7izgiler, iki s\u0131n\u0131f\u0131 ay\u0131ran \u201cideal\u201d s\u0131n\u0131rd\u0131r. Bu b\u00f6l\u00fcm\u00fcn geri kalan\u0131nda, az\u0131nl\u0131k \u00f6rneklerine at\u0131fta bulunurken pozitif, \u00e7o\u011funluk \u00f6rneklerine at\u0131fta bulunurken negatif kullanaca\u011f\u0131z.<\/span><\/p>\n

S\u0131n\u0131fland\u0131r\u0131c\u0131 olarak DVM’leri kullanarak, (a) ve (b)’de dama tahtas\u0131n\u0131n sol iki \u00e7eyre\u011fi aras\u0131ndaki s\u0131n\u0131r bozulmas\u0131n\u0131 iki farkl\u0131 negatif\/pozitif e\u011fitim verisi oran\u0131 alt\u0131nda \u00e7izdik.<\/span><\/p>\n

Negatif \u00f6rneklerin say\u0131s\u0131n\u0131n (yukar\u0131daki \u00e7eyrekte) pozitif \u00f6rneklerin say\u0131s\u0131na (a\u015fa\u011f\u0131daki \u00e7eyrekte) oran\u0131 10:1 oldu\u011funda SVM s\u0131n\u0131f\u0131 s\u0131n\u0131r\u0131n\u0131 g\u00f6sterir. Oran 10.000:1’e y\u00fckseldi\u011finde s\u0131n\u0131r\u0131 g\u00f6sterir. S\u0131n\u0131r, \u015eekil 6(a)’daki s\u0131n\u0131rdan pozitif \u00e7eyre\u011fe do\u011fru \u00e7ok daha e\u011fimlidir, bu da yanl\u0131\u015f negatiflerin daha y\u00fcksek g\u00f6r\u00fclme s\u0131kl\u0131\u011f\u0131na neden olur.<\/span><\/p>\n

Hem sorgu kavram\u0131 \u00f6\u011frenme hem de semantik a\u00e7\u0131klama, dengesiz e\u011fitim s\u0131n\u0131f\u0131 probleminden zarar g\u00f6rebilir, \u00e7\u00fcnk\u00fc hedef s\u0131n\u0131f\u0131n say\u0131s\u0131 di\u011fer s\u0131n\u0131flara g\u00f6re her zaman \u00e7ok fazlad\u0131r (yani, ilgili s\u0131n\u0131f genellikle nadirdir ve alakas\u0131z s\u0131n\u0131f taraf\u0131ndan say\u0131ca fazlad\u0131r).<\/span><\/p>\n

DVM’ler i\u00e7in dengesiz e\u011fitim-veri sorununu \u00e7\u00f6zmek i\u00e7in iki yakla\u015f\u0131m\u0131 ele al\u0131yoruz: algoritmik yakla\u015f\u0131m ve veri i\u015fleme yakla\u015f\u0131m\u0131. Algoritmik yakla\u015f\u0131m, \u00e7ekirdek i\u015flevini veya \u00e7ekirdek matrisini de\u011fi\u015ftirmeye odaklan\u0131r ve veri i\u015fleme yakla\u015f\u0131m\u0131, e\u011fitim verilerini i\u015flemeye odaklan\u0131r.<\/span><\/p>\n

Algoritmik \u00c7\u00f6z\u00fcm<\/span><\/strong><\/h3>\n

Karar sonucunu etkileyen \u00fc\u00e7 parametre tan\u0131mlayabiliriz: b, \u03b1i ve K. Burada, dengesiz e\u011fitim s\u0131n\u0131flar\u0131 sorununu \u00e7\u00f6zmek i\u00e7in tek etkili algoritmik y\u00f6ntemin, K \u00e7ekirde\u011fini veya \u00e7ekirdek matrisini2 Ki,j uyarlamal\u0131 olarak de\u011fi\u015ftirmek oldu\u011funu g\u00f6steriyoruz.\u00a0<\/span><\/p>\n

\u00d6zellik alan\u0131ndaki ay\u0131r\u0131c\u0131 s\u0131n\u0131r y\u00fczeyi etraf\u0131ndaki az\u0131nl\u0131k ve \u00e7o\u011funluk alanlar\u0131n\u0131n uzamsal \u00e7\u00f6z\u00fcn\u00fcrl\u00fc\u011f\u00fcn\u00fc dinamik olarak b\u00fcy\u00fctmek (veya \u00f6l\u00e7eklendirmek) i\u00e7in uyarlanabilir konformal d\u00f6n\u00fc\u015f\u00fcm\u00fc kullanmay\u0131 \u00f6neriyoruz, b\u00f6ylece s\u0131n\u0131r daha az e\u011fri olurken SVM marj\u0131 art\u0131r\u0131labilir.\u00a0<\/span>Konformal d\u00f6n\u00fc\u015f\u00fcm, e\u015fleme i\u015flevi taraf\u0131ndan ind\u00fcklenen Riemann geometrik yap\u0131s\u0131n\u0131 de\u011fi\u015ftirerek \u00e7ekirdek i\u015flevini de\u011fi\u015ftirir.<\/span><\/p>\n

Vekt\u00f6r uzay\u0131 temsili olmayan veriler i\u00e7in (video dizisi verileri gibi), d\u00f6n\u00fc\u015f\u00fcm\u00fc do\u011frudan \u00e7ekirdek matrisine uygular\u0131z. Hem UCI hem de ger\u00e7ek d\u00fcnya g\u00f6r\u00fcnt\u00fc\/video veri k\u00fcmeleri \u00fczerindeki deneysel sonu\u00e7lar\u0131m\u0131z, \u00f6nerilen algoritman\u0131n \u00e7arp\u0131k s\u0131n\u0131r\u0131 d\u00fczeltmede \u00e7ok etkili oldu\u011funu g\u00f6stermektedir.<\/span><\/p>\n

Veri \u0130\u015fleme \u00c7\u00f6z\u00fcm\u00fc<\/span><\/strong><\/h3>\n

Torbalama i\u015flemi, e\u011fitim verilerini bir dizi torbaya alt-\u00f6rnekler, her bir torbay\u0131 e\u011fitir ve nihai s\u0131n\u0131f tahminleri yapmak i\u00e7in torbalar\u0131n kararlar\u0131n\u0131 toplar. Alt \u00f6rneklemeyi biraz farkl\u0131 \u015fekilde ger\u00e7ekle\u015ftiren yar\u0131 torbalamay\u0131 \u00f6neriyoruz.<\/span><\/p>\n

Her torba i\u00e7in, yar\u0131 torbalama yaln\u0131zca \u00e7o\u011funluk s\u0131n\u0131f\u0131n\u0131n alt \u00f6rneklerini al\u0131r, ancak t\u00fcm az\u0131nl\u0131k s\u0131n\u0131f\u0131n\u0131 al\u0131r. Yaln\u0131zca \u00e7o\u011funluk s\u0131n\u0131f\u0131n\u0131n alt \u00f6rneklemesini yaparak, her e\u011fitim \u00e7antas\u0131nda s\u0131n\u0131f dengesini iyile\u015ftiriyoruz. Ayn\u0131 zamanda, \u00e7ok say\u0131da torba kullan\u0131lmas\u0131, alt \u00f6rnekleme i\u015fleminin neden oldu\u011fu varyans\u0131 azalt\u0131r.<\/span><\/p>\n

\n\n
<\/div>\n\n<\/div>\n \n