Scene Buildup From Latent Memory Representations Across Eye Movements.

Area of Science:

• Cognitive Neuroscience
• Visual Perception
• Computational Neuroscience

Background:

• Understanding how the brain constructs scene representations from sequential eye fixations remains a challenge.
• Existing models of visual working memory struggle to account for the detailed yet sparse nature of scene representations needed for behavior.

Purpose of the Study:

• To propose a novel framework for on-line scene representation construction using latent representations.
• To explain how these latent representations support detailed and flexible information retrieval for behavioral goals.

Main Methods:

• Theoretical modeling of latent representations in working memory.
• Discussion of their properties: activity-silent maintenance, large capacity, and energy efficiency.
• Exploration of their link to observable functional connectivity patterns.

Main Results:

• Latent representations accumulate and interact in working memory to form detailed, task-relevant scene representations.
• These representations are sparse, focusing on fixated information, yet allow flexible attentional prioritization.
• Observable as transient functional connectivity patterns resulting from synaptic weight changes.

Conclusions:

• Latent representations offer a viable mechanism for constructing rich scene representations from sequential fixations.
• This framework reconciles the sparse nature of scene representations with the need for detailed information.
• Future research can leverage EEG-eye movement co-registration to observe these latent representations.