Efficient training of backpropagation neural networks

مقدم أطروحة جامعية

Otair, Muhammad Abd Allah Arif

مشرف أطروحة جامعية

Salamah, Walid Khalid

أعضاء اللجنة

Isa, Ghassan Farid
al-Shammari, Husayn Hadi Uwayyid
Abu al-Suud, Salih Mustafa

الجامعة

الأكاديمية العربية للعلوم المالية و المصرفية

الكلية

كلية نظم و تكنولوجيا المعلومات

دولة الجامعة

الأردن

الدرجة العلمية

دكتوراه

تاريخ الدرجة العلمية

2004

الملخص الإنجليزي

Since the discovery of the Backpropagation algorithm, many modified and new algorithms have been proposed for training of feedforward neural networks.

The problem with slow convergence rate has, however, not been solved when the training is on large-scale problems.

There is still a need for more efficient algorithms.

This Ph.D.

thesis describes different approaches to improve the convergence rate.

The target of this thesis is to discover an Optical Backpropagation (OBP) and the stochastic version of this algorithm.

Other important results are the modifying of existing algorithms that use Backpropagation to achieve better results in versions of different parameters selection.

The OBP algorithm proposes a modified error function to reduce the probability that output nodes are near the wrong extreme value of sigmoid activation function.

This is acquired through a strong error signal for the incorrectly saturated output node and a weak error signal for the correctly saturated output node.

The weak error signal for the correctly saturated output node, also, prevents overspecialization of learning for training patterns.

The major contributions of this thesis are the design and utilization of the OBP algorithm for improving the performance of current supervised training algorithms on different experiments.

These improvements include reduction of the training time for some existing supervised training algorithms (such as backpropagation, Backpropagation with Momentum and Delta-Bar-Delta).

Further, the experimental results show that the OBP algorithm converges to a reasonable range of error after a few number of training epochs, making it suitable for dynamic real-time applications.

التخصصات الرئيسية

تكنولوجيا المعلومات وعلم الحاسوب

الموضوعات

• الخوارزميات
• الشبكات العصبية(الحاسبات الإلكترونية)

عدد الصفحات

195

قائمة المحتويات

• APA
• MLA
• AMA

لغة النص

الإنجليزية

نوع البيانات

رسائل جامعية

رقم السجل

BIM-304755