Characterizing the nonlinear structure of shared variability in cortical neuron populations using latent variable...

Area of Science:

• Neuroscience
• Computational Neuroscience
• Systems Neuroscience

Background:

• Sensory neuron responses to repeated stimuli are variable, potentially losing information.
• Shared variability across neurons can preserve stimulus information, depending on its structure.
• Latent variable models analyze neural variability but often assume restrictive linear relationships.

Purpose of the Study:

• Introduce novel nonlinear latent variable models for large-scale neural recordings.
• Characterize shared neural variability and its relationship to sensory encoding.
• Investigate differences in neural population variability across brain regions and states.

Main Methods:

• Developed a general nonlinear latent variable model agnostic to neuron tuning.
• Introduced the Generalized Affine Model for simultaneous stimulus selectivity and latent variable determination.
• Applied models to neural recordings from anesthetized primary visual cortex (V1) and awake macaque prefrontal cortex (PFC).

Main Results:

• Anesthetized V1 neural activity is well-described by a single additive and single multiplicative latent variable (affine model).
• Awake macaque PFC recordings reveal more general nonlinearities for compactly describing population response variability.
• Nonlinear latent variable models effectively capture complex relationships in shared neural variability.

Conclusions:

• Nonlinear latent variable models offer powerful tools for analyzing population neural variability.
• The complexity of neural variability differs across brain regions and experimental conditions (e.g., anesthesia vs. awake).
• A range of analytical methods may be necessary to study diverse neural systems effectively.