Linking global top-down views to first-person views in the brain

Area of Science:

• Computational neuroscience
• Cognitive science
• Artificial intelligence

Background:

• Humans and animals can switch between first-person and top-down spatial perspectives, crucial for navigation and memory.
• The medial temporal lobe and other cortical areas are implicated in these viewpoint transformations.
• Understanding the neural mechanisms underlying these spatial computations is an ongoing challenge.

Purpose of the Study:

• To model how neural systems might perform viewpoint transformations between first-person and top-down spatial frames of reference.
• To investigate the computational principles underlying the brain's ability to switch perspectives.
• To compare model-generated neural activity with experimental recordings.

Main Methods:

• Utilized variational autoencoders (VAEs), a type of deep learning model, in a robot simulation.
• Trained VAEs to reconstruct first-person views from top-down views and vice versa.
• Analyzed latent variables within the VAEs for similarities to neural recordings.

Main Results:

• Many VAE latent variables exhibited neural-like activity, including place-specific and head-direction tuning, and encoding of object distances.
• Reconstructing first-person views from top-down views highlighted place-specific activity.
• Reconstructing top-down views from first-person views emphasized head-direction-specific activity.
• The model demonstrated robustness to perturbations through remapping, without retraining.

Conclusions:

• The computational model provides insights into how the brain might link and transform different spatial viewpoints.
• Findings suggest distinct neural representations are important for different directions of viewpoint transformation.
• This work advances understanding of the neural basis of spatial cognition and navigation.