Multiscale representations of community structures in attractor neural networks

Area of Science:

• Computational neuroscience
• Cognitive science
• Graph theory

Background:

• Cognition depends on the brain's ability to segment hierarchically structured events across multiple scales.
• Event segmentation is thought to rely on the structure of state-transition graphs underlying sequential experiences.
• The precise neural circuit mechanisms for this process remain poorly understood.

Purpose of the Study:

• To propose a novel attractor network model for graph-based hierarchical computation in the brain.
• To elucidate the circuit mechanisms underlying multiscale event segmentation.
• To connect graph theory and attractor dynamics for a biologically plausible model of abstraction.

Main Methods:

• Developed an extended attractor network model termed the Laplacian associative memory.
• Analyzed the model's ability to generate multiscale representations corresponding to graph Laplacian eigenvectors.
• Investigated the role of heterogeneous modulation of inhibitory circuits in regulating representation scale.
• Simulated chunked sequential activity patterns.

Main Results:

• The Laplacian associative memory model generates multiscale representations linked to graph communities.
• These representations mathematically correspond to graph Laplacian eigenvectors, a known segmentation method.
• The model successfully reproduces chunked sequential activity patterns similar to hippocampal theta sequences.
• Demonstrated a connection between graph theory, attractor dynamics, and hierarchical brain computation.

Conclusions:

• The proposed model offers a biologically plausible mechanism for hierarchical event segmentation and abstraction in the brain.
• It bridges computational graph theory with neural attractor dynamics.
• The model provides insights into how the brain represents and processes sequential information at multiple scales.