1. UKM Learning and Research Repository
2. PHD Thesis / Tesis PHD
3. Faculty of Information Science and Technology / Fakulti Teknologi dan Sains Maklumat

Title:	Weight and delay adaptation learning in spiking neural network
Authors:	Abdullah Almasri (P43999)
Supervisor:	Shanorbanun Sahran, Prof. Madya Dr.
Keywords:	Neural network Spiking neural networks Artificial neural networks
Issue Date:	4-Nov-2015
Description:	Spiking neural networks (SNN) are the third generation of Artificial Neural Networks (ANN). The increased interest in such SNN has come from the disputes over the structure and functionality of conventional ANN. SNN is capable of dealing with spike-time based coding. SNN can simulate arbitrary feed-forward sigmoidal neural network; and thus approximate any continuous function, even with a seemingly increased structural complexity. Moreover, it was also proven that neurons which convey information by timing of individual spikes are computationally more powerful than the classical neurons with sigmoidal activation function. Currently, SNN has been applied to various applications in automatic pattern recognition. However, it still lacks good learning algorithm. Although considerable research has been devoted to application of SNN, less attention has been paid to mathematically rigorous analysis of the computational power of SNN. Such analysis will be helpful in understanding the behaviour of computations in complex biological neural systems. The first purpose of the present research is to present an analysis of the existing learning algorithms of SNN for pattern recognition application specifically via mathematical rigorous analysis with critical evaluation on the existing method. Ten datasets have been used from UCI. The contribution of this research is as follows; the noisy problem of the scaling method at the pre-processing learning stage is tackled. Outlining the time window boundary, specifying the spike time range and determining the threshold boundary parameters respectively are carried out for learning in temporal coding SNN for both classification and clustering applications. Different strategies are designed and then applied to find out threshold boundary, namely, threshold average, each pattern threshold range and each pattern threshold Spike Response Function (SRF). Those three parameters namely time window, spike time, and threshold had experimentally been assigned in the previous research. This finding decreases the computational cost and increases SNN learning performance efficiently. The second purpose of the research is specifying the delay rules to control the SRF stability. Here, a new technique is proposed to update weight and delay simultaneously at the learning rule through specifying the delay rules which control the SRF stability at a preparatory processing stage. This is applied only for classification application. This finding proves that more study is needed on updating weight and delay simultaneously at the learning rule in a logic way through studying the main role of SRF on this issue. The third purpose of this research is to propose a new model: SNN weight and delay learning vector quantization (SNN-WD-LVQ), for learning in temporal coding SNN. This model give good performance in terms of stability compared to the existing SNN learning models in mathematical perspective. This model is applied to a supervised learning.,Ph.D.
Pages:	204
Publisher:	UKM, Bangi
Appears in Collections:	Faculty of Information Science and Technology / Fakulti Teknologi dan Sains Maklumat

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.