Multistability analysis for recurrent neural networks with unsaturating piecewise linear transfer functions

Area of Science:

• Computational Neuroscience
• Artificial Neural Networks
• Dynamical Systems Theory

Background:

• Monostable neural networks are computationally limited for complex tasks like decision-making.
• Multistability in neural networks is essential for advanced applications requiring multiple stable states.
• Recurrent neural networks (RNNs) with piecewise linear transfer functions are a key area of study.

Purpose of the Study:

• To analyze the multistability properties of recurrent neural networks with unsaturating piecewise linear transfer functions.
• To establish conditions for boundedness, global attractivity, and complete convergence in these networks.
• To provide theoretical foundations and practical examples for designing computationally capable multistable neural networks.

Main Methods:

• Derivation of conditions for network boundedness based on local inhibition.
• Establishment of conditions and bounds for global attractivity.
• Development of complete convergence conditions using novel energy-like functions.
• Utilization of simulation examples to validate theoretical findings.

Main Results:

• Sufficient conditions for boundedness of multistable networks are derived using local inhibition.
• Conditions and bounds for global attractivity of the network states are established.
• Novel energy-like functions are used to develop conditions for complete convergence.
• Simulations confirm the theoretical results, demonstrating the network's stability properties.

Conclusions:

• The study provides a comprehensive analysis of multistability in a specific class of RNNs.
• The derived conditions enhance the understanding and design of neural networks for complex computational tasks.
• This work contributes to the development of more robust and capable artificial intelligence systems.