[Paper Review] A model of cortical cognitive function using hierarchical interactions of gating matrices in internal agents coding relational representations

Michael E. Hasselmo|arXiv (Cornell University)|Sep 21, 2018

Neural dynamics and brain function68 references3 citations

TL;DR

This paper proposes a neural network model of cortical cognition that uses hierarchical gating matrices within internal agents to detect and predict relational structures in sensory inputs. By modeling dynamic relations and affine transformations, the framework successfully predicts sequences and solves abstract reasoning tasks like the Ravens Progressive Matrices, demonstrating a biologically plausible mechanism for flexible, predictive cognition in the neocortex.

ABSTRACT

Flexible cognition requires the ability to rapidly detect systematic functions of variables and guide future behavior based on predictions. The model described here proposes a potential framework for patterns of neural activity to detect systematic functions and relations between components of sensory input and apply them in a predictive manner. This model includes multiple internal gating agents that operate within the state space of neural activity, in analogy to external agents behaving in the external environment. The multiple internal gating agents represent patterns of neural activity that detect and gate patterns of matrix connectivity representing the relations between different neural populations. The patterns of gating matrix connectivity represent functions that can be used to predict future components of a series of sensory inputs or the relationship between different features of a static sensory stimulus. The model is applied to the prediction of dynamical trajectories, the internal relationship between features of different sensory stimuli and to the prediction of affine transformations that could be useful for solving cognitive tasks such as the Ravens progressive matrices task.

Motivation & Objective

• To develop a neurobiologically plausible framework for flexible, predictive cognition in the neocortex.
• To explain how neural circuits can detect and represent systematic relations between sensory features and temporal sequences.
• To model internal cognitive processes as dynamic interactions among gating agents that control connectivity patterns.
• To apply the model to predictive tasks such as sequence prediction and transformation inference in abstract reasoning problems.
• To demonstrate that hierarchical gating of relational representations can support high-level cognitive functions like those seen in human problem-solving.

Proposed method

• The model employs internal agents that represent neural populations and use gating matrices to control connectivity between them.
• Gating matrices are dynamically modulated to represent functional relations between neural populations, enabling prediction of future states.
• The framework uses column vectors for state representations and matrix operations to model hierarchical interactions.
• Relations between sensory inputs are encoded as patterns of connectivity that are gated based on contextual and predictive signals.
• The model applies to both temporal sequences and static stimulus relationships, including affine transformations.
• Predictive performance is evaluated using tasks such as sequence continuation and solving the Ravens Progressive Matrices.

Experimental results

Research questions

• RQ1 How can neural circuits detect and represent systematic relations between components of sensory input?
• RQ2 What mechanisms allow the brain to predict future states based on current sensory inputs and relational structures?
• RQ3 How can hierarchical interactions of gating matrices support flexible, context-dependent cognitive processing?
• RQ4 Can a biologically plausible neural model solve abstract reasoning tasks like the Ravens Progressive Matrices?
• RQ5 How do dynamic gating mechanisms enable the prediction of transformations in relational structures?

Key findings

• The model successfully predicts dynamical trajectories in sequential sensory inputs using hierarchical gating of relational representations.
• It accurately models internal relationships between features of static sensory stimuli through learned connectivity patterns.
• The framework can infer and apply affine transformations, which are essential for solving abstract reasoning problems such as those in the Ravens Progressive Matrices.
• The use of column vectors and simplified notation improves clarity and consistency in modeling neural state transitions.
• The model demonstrates that internal gating agents can simulate cognitive functions like prediction and relational reasoning using biologically plausible mechanisms.
• The revised version (v2) improves mathematical clarity and corrects typographical errors without altering core results.