Enhancing the performance of the back propagation algorithm for deep neural network

العناوين الأخرى

تحسين أداء خوارزمية الانتشار الخلفي في الخلايا العصبية الصناعية

مقدم أطروحة جامعية

Surakhi, Ula Muhammad

مشرف أطروحة جامعية

Salamah, Walid A.

أعضاء اللجنة

Umar, Khamis
Ghnimat, Rawan
al-Dawud, Ali Sad

الجامعة

جامعة الأميرة سمية للتكنولوجيا

الكلية

كلية الملك الحسين لعلوم الحوسبة

القسم الأكاديمي

قسم علم الحاسوب

دولة الجامعة

الأردن

الدرجة العلمية

ماجستير

تاريخ الدرجة العلمية

2014

الملخص الإنجليزي

One major problem encountered by researchers in improving the performance of the Backpropagation algorithm is the slow convergence and convergence to the local minima.

Many modified and new versions on the Backpropagation since it has been used and launched.

Those have proposed a careful selection on the initial weights and biases, learning rate, momentum, network topology and activation function.

As the algorithm is widely used in solving many real problems in the world there is still a sever need for more efficient modification.

This thesis, in fact, will describe a new approach to enhance the performance of training multi-layer neural networks and deep neural networks with more than one hidden layer by using a new algorithm which is called the Extended Optical Backpropagation (EOBP).

A new error function has been adopted to replace the error function used in Optical Backpropagation OBP algorithm which gives a rapid reaction to changes in the weights value by increasing the training speed with less number of iterations and without loss of learn-ability.

Experiments have been conducted to compare and to evaluate the convergence behavior of these training algorithms with two training problems: XOR and the Iris plant classification.

The results showed that the proposed algorithm converges to a reasonable range of error after a few numbers of training epochs

التخصصات الرئيسية

الرياضيات

الموضوعات

• الخوارزميات
• الشبكات العصبية(الحاسبات الإلكترونية)

عدد الصفحات

67

قائمة المحتويات

• APA
• MLA
• AMA

لغة النص

الإنجليزية

نوع البيانات

رسائل جامعية

رقم السجل

BIM-535254