Network structure and input integration in competing firing rate models for decision-making.

Area of Science:

• Computational Neuroscience
• Decision Making Models
• Neural Network Dynamics

Background:

• Mammalian decision-making, especially with multiple options, is complex and not fully understood.
• Existing models often focus on two-option choices, limiting insights into real-world scenarios.

Purpose of the Study:

• To develop and analyze a scalable mechanistic network model for decision-making with numerous alternatives.
• To investigate the impact of network structure and input integration on decision dynamics.

Main Methods:

• Formulated a mechanistic network model for decision-making.
• Analyzed model dynamics for fully-connected and sparse network topologies (random, regular, small-world).
• Characterized fixed points and stability, comparing sigmoidal and binary input integration functions.

Main Results:

• Sigmoidal transfer functions are evolutionarily advantageous over binary gain for input integration.
• Smaller steepness in sigmoidal functions increases speed but decreases accuracy; however, they enhance robustness against noise and connection degradation.
• A stable parameter regime for sigmoidal gain ensures high accuracy across diverse tasks and network structures, meeting energetic constraints.

Conclusions:

• Neural system architecture is potentially optimized for economical, reliable, and advantageous decision-making.
• The model demonstrates how network topology and function (e.g., sigmoidal transfer) contribute to robust multi-alternative decision processes.