A neural network model of when to retrieve and encode episodic memories

Area of Science:

• Cognitive Neuroscience
• Artificial Intelligence
• Computational Psychiatry

Background:

• Human studies indicate selective encoding and retrieval of episodic memories.
• Understanding the mechanisms and benefits of memory selectivity is crucial.

Purpose of the Study:

• To model human episodic memory selectivity using a neural network.
• To investigate factors influencing selective memory retrieval and encoding.
• To provide a resource-rational explanation for memory selectivity.

Main Methods:

• Trained a memory-augmented neural network on an environment with recurring situations.
• Analyzed network's retrieval and encoding strategies based on state uncertainty.
• Evaluated the impact of selective encoding (end vs. mid-event) on prediction performance.

Main Results:

• The network learned to retrieve memories selectively, influenced by uncertainty about future states.
• Selective encoding at the end of an event improved subsequent prediction accuracy.
• Benefits of selectivity were linked to reducing the retrieval of irrelevant memories.

Conclusions:

• Selective episodic memory encoding and retrieval are resource-rational strategies.
• Modeling supports the role of uncertainty and event structure in memory selectivity.
• Findings generate testable predictions for future behavioral and neuroimaging research.