A Feed-Forward Neural Network for Increasing the Hopfield-Network Storage Capacity

Area of Science:

• Neuroscience
• Computational Neuroscience
• Artificial Intelligence

Background:

• Pattern separation in the hippocampal dentate gyrus (DG) is crucial for memory formation.
• Existing theories of DG pattern separation, like 'expansion recoding,' face neurophysiological challenges.
• There is a need for computational models that better reflect DG structure and function.

Purpose of the Study:

• To propose a novel feed-forward neural network inspired by the DG's structure and neural oscillations.
• To increase the storage capacity of Hopfield networks using bio-inspired principles.
• To investigate computational principles of pattern separation in the DG.

Main Methods:

• Developed a novel feed-forward neural network incorporating DG structure and neural oscillatory analysis.
• Established a mouse model of environmental enrichment to study DG function.
• Utilized neural oscillatory analysis to derive a computational model of the DG.
• Created a new algorithm based on Hebbian learning and neural oscillation coupling for network training.

Main Results:

• The proposed bio-inspired network significantly expands the storage capacity of Hopfield networks.
• Achieved more effective pattern separation compared to standard models.
• Storage capacity increased from 0.13 to 0.32 (at 10% pattern overlap).

Conclusions:

• The novel neural network effectively models DG functions for enhanced pattern separation.
• Bio-inspired computational models can improve memory storage capacity.
• This approach offers a new avenue for understanding and replicating hippocampal memory mechanisms.