A neural network model for online one-shot storage of pattern sequences

Area of Science:

• Neuroscience
• Computational Neuroscience
• Cognitive Science

Background:

• The hippocampus is crucial for memory formation, particularly sequential memories.
• Existing models often require extensive consolidation periods for sequence learning.
• The CRISP theory provides a framework for understanding content representation, intrinsic sequences, and pattern completion.

Purpose of the Study:

• To present a novel computational model of the hippocampus capable of online one-shot storage of pattern sequences.
• To investigate the role of CA3 and DG subregions in sequence learning and retrieval.
• To implement a unified learning rule with a forgetting mechanism for continuous pattern storage.

Main Methods:

• Developed a computational model based on the CRISP theory.
• Utilized hetero-association in CA3 instead of direct sequence storage.
• Applied Hebbian descent learning rule with a forgetting mechanism to all plastic synapses.
• Tested the model with artificial sequences, handwritten digits, and natural images.

Main Results:

• The model achieves online one-shot storage of pattern sequences without consolidation.
• A single cue pattern reliably triggers sequence retrieval, even with noisy or incomplete cues.
• Pattern separation in DG is essential for correlated patterns within sequences.
• The model demonstrates self-improvement through a replay-like process.

Conclusions:

• The proposed hippocampus model offers an efficient mechanism for rapid sequence learning and retrieval.
• The model's architecture and learning rule support online storage and consolidation through replay.
• This work advances computational models of memory by incorporating efficient sequence processing and self-improvement.