A Model of Semantic Completion in Generative Episodic Memory.

Area of Science:

• Cognitive Neuroscience
• Computational Neuroscience
• Artificial Intelligence

Background:

• Episodic memory is increasingly viewed as generative, yet computational models often focus on storage.
• Existing models do not fully capture how incomplete memory traces are reconstructed.
• The interplay between episodic memory and semantic information during recall is not well understood computationally.

Purpose of the Study:

• To present a novel computational model of the generative aspects of episodic memory.
• To investigate the role of semantic completion in reconstructing incomplete memory traces.
• To explore how context congruence and attention influence episodic memory recall.

Main Methods:

• Developed a generative model combining Vector-Quantized Variational Autoencoder (VQ-VAE) and Pixel Convolutional Neural Network (PixelCNN).
• Utilized MNIST digits and fashion items with varied backgrounds to represent episodic memory traces and context.
• Modeled an episodic memory experiment to test recall under different contextual conditions.

Main Results:

• The model successfully completed missing parts of memory traces, generating plausible images and generalizing to novel data.
• Semantic completion and compression led to reduced memory requirements and increased robustness to noise.
• Simulated experiments confirmed that semantically congruent contexts improve recall, attention enhances accuracy, and errors often retain semantic congruence.

Conclusions:

• The hippocampus-neocortex model provides a framework for understanding episodic memory as a generative process.
• Semantic completion is a key mechanism for reconstructing past experiences from incomplete memory traces.
• The findings highlight the crucial role of semantic information and attention in accurate episodic memory recall.