|
KMID : 0667720000370000413
|
|
Report Natlonal Institute of Health
2000 Volume.37 No. 0 p.413 ~ p.414
|
|
|
Effect of Synaptic Loss on Learning and Memory Artificial Neural Networks
|
|
김창근/Kim, C. K.
윤선희/백은경/고인송/Yoon, S. H./Paik, E. K./Koh, I. S.
|
|
|
Abstract
|
|
|
Objective : In a Backpropagation neural network(BPNN), we studied the effect of synaptic loss on learning and memory as an attempt to develop a cognitive impairment model for Alzheimer´s disease(AD).
Methods : A fully-connected three-layered feedforward Backpropagation neural network model was developed using Visual C++ 6.0 and Matlab The supervised training method of minimum error learning was used. After training, in order to make an AD model, synaptic deletion has been carried out either randomly or weak-synapse-firstly. In order to evaluate memory performance, the recognition rate was examined. In order to simulate the AD patient´s disability to learn new things, the synaptic sites were excluded randomly to a desired proportion in an intact state but not yet trained BPNN, then the BPNN was trained.
Results : The decline pattern of memory performance varies with the synaptic deletion methods. When random deletion method is applied, the decline shows a linear pattern, whereas when the weak-synapse-first deletion method is used, it shows a plateau and gradually decreasing pattern. In the simulation, synaptic compensation tends to bring delayed memory impairment. Thus synaptic compensation observed in the brain of AD is probably for minimizing memory impairment due to synaptic loss by increasing the remaining synapse size. Learning disability in AD was also demonstrated in the computer model. When the amount of synaptc loss is very high, the network never learn correctly as observed in AD patients.
Conclusions : Memory impairment began to appear after a certain extent of synaptic loss in the weak-synapse-first deletion. This fact might explain delayed development of AD symptom even after synaptic loss has been started. We assume that weak synapse probably dies first, because it is structurally less stable. We also demonstrated that synaptic compensation could slow down the process of memory impairment in AD. We do not insist, however, that an artificial neural network is an acceptable brain model, and that the damaged network is an AD model. We only tried to explain AD symptom development using BPNN and test the possibility of BPNN as a model for AD symptom development. The simulation results give us insights to understand some aspects of AD symptom development and progression.
|
|
|
KEYWORD
|
|
|
|
|
|
FullTexts / Linksout information
|
|
|
|
|
|
Listed journal information
|
|
|
|
|