Positive Neural Networks in Discrete Time Implement Monotone-Regular Behaviors

Area of Science:

• Computational Neuroscience
• Theoretical Computer Science

Background:

• Neural networks are fundamental computational models.
• Understanding their expressive power is crucial for AI development.
• Positive neural networks offer a constrained yet potentially powerful framework.

Purpose of the Study:

• To investigate the expressive capabilities of positive neural networks.
• To determine the class of behaviors implementable by these networks.
• To analyze the role of time delay in their functionality.

Main Methods:

• Analysis of positive neural network models in discrete time.
• Characterization of implementable behaviors based on connection weights.
• Investigation of monotone-regular behaviors and their relation to regular languages.

Main Results:

• Positive neural networks capture monotone-regular behaviors.
• All such behaviors are implementable with a one-time-unit delay.
• Some behaviors are implementable with zero delay, while others are not.

Conclusions:

• Positive neural networks possess significant expressive power, equivalent to monotone-regular behaviors.
• Time delay is a critical factor influencing the implementation capabilities of these networks.
• The study provides insights into the fundamental computational limits and possibilities of constrained neural architectures.