Learning to represent continuous variables in heterogeneous neural networks.

Area of Science:

• Computational Neuroscience
• Neural Network Theory
• Animal Behavior

Background:

• Animals continuously monitor variables like position and head direction.
• Manifold attractor networks, with persistent neuronal states, are a leading model for this ability.
• Existing models often assume symmetric synaptic connectivity, which contradicts experimental observations.

Purpose of the Study:

• To develop a theory for manifold attractors in trained neural networks without assuming unrealistic symmetry.
• To predict the effects of representational asymmetries and connectivity heterogeneity on manifold formation and stability.
• To link functional properties of brain manifold attractors to asymmetries in connectivity and low-dimensional representations.

Main Methods:

• Developed a new theory for manifold attractors in trained neural networks.
• Utilized theoretical modeling to analyze the impact of asymmetries and heterogeneity.
• Investigated how these factors influence manifold formation, network response, and stability.

Main Results:

• Demonstrated that manifold attractors can form without symmetric synaptic connectivity.
• Predicted how representational asymmetries and heterogeneous connectivity shape manifold properties.
• Identified mechanisms potentially leading to manifold destabilization.

Conclusions:

• The functional properties of brain manifold attractors can be understood by considering asymmetries.
• Overlooked asymmetries in connectivity and representation are crucial for network function.
• This theory offers a more biologically plausible explanation for neuronal monitoring.