Related Experiment Videos
A comparative study of two modeling approaches in neural networks
Zong-Ben Xu1, Hong Qiao, Jigen Peng
1Institute for Information and System Sciences, Xi'an Jiaotong University, Xi'an, China. zbxu@mail.xjtu.edu.cn
Summary
This study compares static and local field neural network models, revealing a key stability invariance property. This finding supports a unified stability analysis methodology for diverse neural network models.
Area of Science:
- Computational Neuroscience
- Artificial Neural Networks
- Theoretical Neuroscience
Background:
- Two fundamental approaches to neural network research are neuron state modeling and local field modeling.
- These approaches lead to static neural network models and local field neural network models, respectively.
Purpose of the Study:
- To theoretically compare static and local field neural network models.
- To investigate properties such as trajectory transformation, equilibrium correspondence, manifold properties, convergence, and stability.
- To establish a foundation for a unified stability analysis methodology.
Main Methods:
- Theoretical comparison of static and local field neural network models.
- Analysis of properties including trajectory transformation, equilibrium correspondence, and stability in various senses.
- Mathematical derivation of stability invariance property.
Main Results:
- A significant stability invariance property was identified between the two models.
- The stability of a static model is equivalent to a subsystem of the local field model under specific conditions.
- This invariance holds across different senses of stability.
Conclusions:
- The identified stability invariance property provides a theoretical basis for cross-model stability analysis.
- This facilitates a unified methodology for analyzing the stability of various neural network models.
- The findings contribute to a deeper understanding of neural network dynamics and stability.