Minicolumn-Based Episodic Memory Model With Spiking Neurons, Dendrites and Delays

Area of Science:

• Computational Neuroscience
• Cognitive Science
• Artificial Intelligence

Background:

• Episodic memory is crucial for cognitive function, yet the temporal organization of neuronal activity during its processes remains unclear.
• Existing models lack a comprehensive understanding of how sequential information is encoded, stored, and retrieved in the brain.

Purpose of the Study:

• To propose a novel Bionic Spiking Temporal Memory (BSTM) model integrating hippocampus structure and spiking neural networks (SNNs).
• To explore the encoding, formation, and retrieval mechanisms of episodic memory using biologically inspired principles.

Main Methods:

• Developed a BSTM model utilizing bionic spiking neurons with biological characteristics (columnar/dendritic structures, spike transmission).
• Employed spike-timing-dependent plasticity (STDP) and a minicolumn selection algorithm for encoding memory items.
• Implemented sequential and local/global retrieval algorithms for context storage and sequence recall.

Main Results:

• The BSTM model successfully encoded and retrieved episodic memory sequences, demonstrating multisentence and multitime step prediction capabilities.
• Experiments using the Children's Book Test (CBT) dataset showed the BSTM model achieved higher accuracy and robustness compared to other algorithms.
• Performance was validated across various settings, including adjustments to model parameters like minicolumns, neurons, and sequence complexity.

Conclusions:

• The BSTM model provides a biologically plausible framework for understanding episodic memory functions.
• This SNN-based approach offers a robust and accurate method for simulating complex memory processes.
• The findings advance the understanding of neural mechanisms underlying memory encoding and retrieval.