Thesis (M.Sc.) - Cairo University - Faculty Of Engineering - Department Of Biomedical Engineering

Artificial Neural Networks (ANN) has proved to be a useful tool for information processingThe problem with conventional ANN is that it assumes that the synapses are staticIn human nervous system , however , synapses are dynamicThese dynamics arise from the influence of pre - synaptic mechanisms on the release from an axon terminalThese mechanisms are facilitation and pre - synaptic feedback inhibition which are fundamental features of biological neuronsConsequently , the probability of release becomes a function of the temporal pattern of action potential occurrenceHence , the strength of a given synapse varies upon the arrival of each action potential invading the terminal regionWe developed an implementation for a stochastic model as computational tool for pattern recognitionWe use the computational capacity of the dynamic synapses for transforming the temporal pattern of spikes into a spatial - temporal pattern of synaptic eventsIn this implementation , the system is learned to extract a short statistically significant sub - pattern in the spike trainSuch a capability for extracting invariant temporal features allows the transmission of information contained within noisy and variable spike trainsThe computational performance is evaluated qualitatively and quantitatively in comparison with the traditional artificial neural networks and a basic version of the stochastic modelAfter training , the network generates high correlated signals when the input signals are similar in there temporal and statistical featuresThe network is tested also for different temporal but same statistical features in input signalsAnother configuration is used for speech signal recognitionThe speech recognition was to differ between two spoken words without any control on the ambient noiseThe results are ensuring that despite how small is the network , the performance is higher than another approach utilizing the same concept of dynamic synapses but in different integrationThe results ensure the enhanced performance of the proposed approach over the other mentioned techniquesMoreover , the embedded encoding within the neural configurations is analyzed by the relative entropy , Kullback - Leibler distanceThis measurement ensured that the network is dealing with the different input signals with statistically dependant manner

Issued also as CD