Publication: Çoklu adaptif hibrit sinir ağları
Abstract
v ÖZET ÇOKLU ADAPTİF HİBRİT SİNİR AĞLARI Evrişimsel sinir ağları, özellikle bilgisayarla görme problemlerini oldukça başarılı bir şekilde çözmekte ve her geçen gün de yeni gelişen mimariler ile daha iyi sonuçlar vermeye devam etmektedir. Hem geliştirilen yeni mimariler ile gelişmekte ama bu mimarilerin de birbirlerinin karışımıyla oluşturulan hibrit modeller de bazı problemlere oldukça iyi sonuçlar da vermektedir. Bu çalışmada farklı mimarilerden sentezlenen modüllerin karışımıyla bir hibrit model oluşturulmuştur. Ancak farklı mimariler kolektif öğrenmede olduğu gibi her modüle eğitilebilir ağırlık katsayısı eklenmektedir. Böylelikle, düşük ağırlıklı modüllerin elenerek model için en uygun modüllerin seçilmesi hedeflenmiştir. Bu modüller farklı ağ mimarilerinde kullanılan yapılardan oluşmaktadır. Bu modüllerin toplanıp seçilmesinin yapıldığı bir katmandan oluşmaktadır. Başlangıçta eşit olan bu ağırlıklar bir eğitim sürecinden sonra incelenerek, önceden belirlenmiş bir eşik değerin altında ise o modül modelden çıkarılmış veya bir başka önceden belirlenmiş eşik değerinin üstünde ise de o modül korundu ve diğer modüller modelden çıkarılmıştır. Bu öğrenme yönetimi kolektif öğrenme yapısına benzemektedir ancak bu çalışmada katsayılar sonradan modelin seçilmesini ya da elenmesini sağlayacaktır. Modüller seçildikten sonra model eğitilmeye devam edilerek başarımı koruyarak model küçültülmüştür. Farklı modüller kullanılarak dinamik bir hibrit modül oluşturulması amaçlanmıştır. Hata ayıklama için kullanılan MNIST ve Fashion MNIST veri kümelerinde sırasıyla %99,09 ve %99,02 doğruluk değerleri elde edilmiştir. Kalite testlerinde kullanılan CIFAR10 veri kümesinde %87,86 doğruluğa ulaşmışken, CIFAR100 veri kümesinde de %54,11 doğruluk sonuçlarına ulaşılmıştır.
Convolutional neural networks, especially with computer vision problems quite successfully and continue to give better results with newly developing architectures day by day. It develops with the developed architectures, but hybrid models created by mixing these architectures with each other also gave very good results for some problems. In this study, a hybrid model was created with a mixture of modules used from different architectures. However, as in collective learning of different architectures, a trainable weight coefficient is added to each module. Thus, it was aimed to select the most suitable modules for the model by eliminating the low-weight modules. These modules consist of structures used in different network architectures. It consists of a selection layer where these modules are collected and selected. These initially equal weights were examined after a training process, and if it was below a predetermined threshold, that module was removed from the model, or if it was above another predetermined threshold, that module was retained, and other modules were excluded from the model. This learning management is similar to the collective learning structure, but in this study, the coefficients will enable the model to be selected or eliminated later. After the modules were eliminated, the model continued to be trained and the model was shrunk while maintaining its performance. It is aimed to create a dynamic hybrid module by using different modules. MNIST and Fashion MNIST datasets used in debugging tests reached 99.09 % and 99.02 % accuracy respectively. CIFAR10 and CIFAR100 datasets used in benchmark tests also reached 87.86 % and 54.11 % accuracy respectively.
Convolutional neural networks, especially with computer vision problems quite successfully and continue to give better results with newly developing architectures day by day. It develops with the developed architectures, but hybrid models created by mixing these architectures with each other also gave very good results for some problems. In this study, a hybrid model was created with a mixture of modules used from different architectures. However, as in collective learning of different architectures, a trainable weight coefficient is added to each module. Thus, it was aimed to select the most suitable modules for the model by eliminating the low-weight modules. These modules consist of structures used in different network architectures. It consists of a selection layer where these modules are collected and selected. These initially equal weights were examined after a training process, and if it was below a predetermined threshold, that module was removed from the model, or if it was above another predetermined threshold, that module was retained, and other modules were excluded from the model. This learning management is similar to the collective learning structure, but in this study, the coefficients will enable the model to be selected or eliminated later. After the modules were eliminated, the model continued to be trained and the model was shrunk while maintaining its performance. It is aimed to create a dynamic hybrid module by using different modules. MNIST and Fashion MNIST datasets used in debugging tests reached 99.09 % and 99.02 % accuracy respectively. CIFAR10 and CIFAR100 datasets used in benchmark tests also reached 87.86 % and 54.11 % accuracy respectively.