Generalized stability and associative memory of delayed recurrent neural networks with variable external input

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Dynamical Systems

Background:

• Delayed recurrent neural networks (DRNNs) are crucial for complex information processing.
• Understanding their stability and memory capacity is essential for advanced applications.
• Existing models often face limitations in handling variable inputs and maximizing storage.

Purpose of the Study:

• To investigate the generalized stability of DRNNs with variable external inputs.
• To analyze the coexistence and stability of multiple equilibrium points.
• To design a high-capacity associative memory system.

Main Methods:

• Application of the comparison principle to establish monostability.
• Extension of stability concepts to DRNNs with variable inputs.
• Modification of activation functions to increase stable equilibrium points.
• Derivation of sufficient conditions for generalized stability.

Main Results:

• Monostability of normal differentiable systems established and extended to DRNNs.
• Analysis of multiple equilibrium points and enhancement of storage capacity.
• Development of conditions for generalized stability, encompassing exponential stability.
• Successful design of a high-capacity associative memory based on stable bipolar patterns.

Conclusions:

• The study provides a theoretical framework for analyzing and enhancing the stability and memory of DRNNs.
• The proposed methods significantly increase the storage capacity of associative memory.
• Numerical examples validate the theoretical findings and the practical design of the associative memory.