Recurrent predictive coding models for associative memory employing covariance learning.

Area of Science:

• Computational Neuroscience
• Theoretical Neuroscience
• Neuroscience of Memory

Background:

• The hippocampus is crucial for associative memory (AM) tasks.
• Predictive coding theories suggest a unitary account for hippocampal AM and predictive activities.
• Existing hierarchical predictive models lack recurrent connections vital for hippocampal CA3 function.

Purpose of the Study:

• To develop biologically plausible computational models for hippocampal AM.
• To address limitations of previous predictive coding (PC) models in learning covariance.
• To integrate predictive coding with implicit covariance learning in hippocampal network models.

Main Methods:

• Proposed alternative PC networks that implicitly learn covariance information.
• Utilized dendritic structures for encoding prediction errors.
• Performed analytical derivations to demonstrate model equivalence and stability.

Main Results:

• The proposed models perform AM plausibly and without numerical instability.
• Analytical results confirm equivalence to explicit covariance-learning PC models.
• Models can be integrated with hierarchical PC networks for hippocampo-neocortical interaction modeling.

Conclusions:

• The developed models offer a biologically plausible computational mechanism for hippocampal memory.
• These models successfully combine predictive coding with implicit covariance learning.
• The findings provide insights into hippocampal memory formation and recall processes.