Using Generative Models of Naturalistic Scenes to Sample Neural Population Tuning Manifolds.

Area of Science:

• Neuroscience
• Computational Vision
• Machine Learning

Background:

• Traditional visual stimuli (e.g., Gabor patches) are limited for studying complex neural responses.
• Naturalistic image models offer more complex stimuli but pose challenges in mapping to neural activity due to high dimensionality.
• Recording neural populations generates high-dimensional response data, exacerbating the dimensionality challenge.

Purpose of the Study:

• To develop a closed-loop method for efficiently characterizing neural population manifolds in high-dimensional stimulus spaces.
• To investigate the utility of deep generative models for understanding sensory coding in the visual system.

Main Methods:

• A closed-loop experimental design was employed.
• Stimuli were generated using a deep neural network.
• Neural responses guided the iterative generation of stimuli to predict relationships between the model's latent space and neural activity.

Main Results:

• Deep generative image models effectively mapped high-dimensional stimulus spaces to neural population activity.
• Latent variables from the deep generative model showed stronger linear relationships with neural activity compared to other compression methods.
• The proposed method efficiently characterized high-dimensional tuning manifolds.

Conclusions:

• Deep generative models provide a powerful tool for efficient characterization of neural population manifolds.
• This approach advances our understanding of sensory coding in the visual system, particularly for complex stimuli.
• The method addresses the
• curse of dimensionality
• in neural data analysis.