Distributed Representations for Cognitive Control in Frontal Medial Cortex.

Area of Science:

• Computational neuroscience
• Cognitive neuroscience
• Artificial intelligence

Background:

• Neural networks balance modularity (hierarchical representations) and distributed structure (flexible training).
• Understanding these structures is key for modeling human sequential behavior and cognitive control.
• Previous research suggests the anterior cingulate cortex (ACC) encodes task progress.

Purpose of the Study:

• To investigate the competing demands of modularity and distributed structure in artificial neural networks for human sequential behavior.
• To compare recurrent neural network models with and without explicit hierarchical structures.
• To determine which model best explains functional magnetic resonance imaging (fMRI) data from human participants.

Main Methods:

• Compared properties of several recurrent neural network models, including hierarchical and 'flat' architectures.
• Utilized representational similarity analysis to compare model outputs with fMRI data.
• Investigated models incorporating 'wiring costs' to induce hierarchical organization.

Main Results:

• Explicit hierarchical structure alone did not significantly outperform a 'flat' model in general performance.
• Hierarchy facilitated cognitive control for nonroutine behaviors and under computational stress.
• A model with wiring costs, creating a hierarchical representational gradient, best matched human fMRI data.

Conclusions:

• Hierarchical organization in artificial neural networks supports cognitive control and nonroutine tasks.
• The anterior cingulate cortex (ACC) exhibits a rostro-caudal gradient of abstraction in representing sequential task context.
• Findings offer insights into the ACC's role in motivation and cognitive control through distributed representations.