Related Experiment Videos
An efficient training algorithm for dynamic synapse neural networks using trust region methods
Hassan H Namarvar1, Theodore W Berger
1Department of Biomedical Engineering, University of Southern California, OHE-500, Los Angeles, CA 90089-1451, USA. heidarin@usc.edu
Summary
We developed a dynamic synapse neural network model using averaged neuron activity. Our trust region optimization method trains large-scale networks more effectively than genetic algorithms.
Area of Science:
- Computational Neuroscience
- Artificial Intelligence
Background:
- Traditional neural networks often lack dynamic synaptic plasticity.
- Modeling large-scale neural networks presents significant computational challenges.
Purpose of the Study:
- To formulate a dynamic synapse neural network model based on local neuronal population activity.
- To introduce and evaluate a novel trust region nonlinear optimization approach for training these networks.
Main Methods:
- Formulation of a dynamic synapse neural network model.
- Application of trust region nonlinear optimization for network training.
- Comparative analysis against genetic algorithms for large-scale network optimization.
Main Results:
- The proposed trust region optimization effectively trains dynamic synapse neural networks.
- Demonstrated superior performance of the new learning method compared to genetic algorithms on large-scale networks.
Conclusions:
- The trust region nonlinear optimization approach offers an effective and efficient method for training dynamic synapse neural networks.
- This formulation advances the development of biologically plausible and scalable artificial neural networks.