Optimal forgetting: Semantic compression of episodic memories

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Artificial Intelligence

Background:

• Human memory exhibits systematic distortions linked to resource limitations.
• Traditional lossy compression models fail to replicate human memory distortions.
• A novel framework is needed to explain the qualitative nature of memory errors.

Purpose of the Study:

• To propose a semantic compression framework for explaining human memory distortions.
• To demonstrate how generative models can simulate memory compression.
• To unify diverse memory phenomena under a single theoretical model.

Main Methods:

• Formalizing memory resource constraints within a lossy compression framework.
• Utilizing deep generative models to approximate environmental generative models.
• Testing the framework on datasets including chess games, natural text, and sketches.

Main Results:

• The semantic compression model successfully replicates human memory distortions.
• The model accounts for phenomena such as domain expertise effects and gist-based recall.
• Demonstrated effects of context and delayed recall within the compression framework.

Conclusions:

• Semantic compression, driven by environment-adapted generative models, offers a unified explanation for memory distortions.
• Deep generative models provide a powerful tool for simulating and understanding these memory processes.
• This framework advances our understanding of memory's adaptive, rather than purely flawed, nature.