Neural modeling of episodic memory: encoding, retrieval, and forgetting.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence

Background:

• Episodic memory systems are crucial for learning and recalling event sequences.
• Existing models often struggle with noisy input and efficient memory management.

Purpose of the Study:

• To develop a novel neural model for learning and recalling episodic traces from continuous sensory input.
• To enhance robustness and control memory consumption in artificial memory systems.

Main Methods:

• Utilized a fusion adaptive resonance theory (ART) network to extract key events and spatio-temporal relations.
• Implemented a dynamic cognitive node creation and a continuous parallel memory search procedure.
• Integrated a gradual forgetting mechanism to manage memory over time.

Main Results:

• The model demonstrated robust encoding and recall of events and episodes, even with incomplete and noisy cues.
• The forgetting mechanism enhanced performance in noisy environments.
• Achieved higher tolerance to noise and errors in retrieval cues compared to prior spatio-temporal memory models.

Conclusions:

• The proposed fusion ART-based episodic memory model offers a robust and efficient solution for spatio-temporal learning.
• The dynamic node creation and forgetting mechanisms contribute to high performance and memory control.
• This model advances the capabilities of artificial systems in handling complex, real-world sensory data.