Dynamical flexible inference of nonlinear latent structures in neural population activity

Area of Science:

• Computational neuroscience
• Machine learning for neuroscience
• Neural dynamics modeling

Background:

• Neural population activity exhibits complex spatiotemporal dynamics, often obscured by noise.
• Existing models struggle to capture nonlinear dynamics and allow flexible inference (causal, non-causal, handling missing data).

Approach:

• Developed DFINE (Dynamic and Manifold Factor Inference Network), a novel neural network architecture.
• DFINE separates dynamic and manifold latent factors for tractable modeling of nonlinear structures.
• Enables flexible inference across various conditions, including missing neural observations.

Key Points:

• DFINE demonstrates superior performance in predicting neural activity and behavior compared to prior models.
• The model effectively captures the underlying latent neural manifold structure.
• Achieves flexible nonlinear inference across diverse brain regions and behavioral tasks.

Conclusions:

• DFINE offers a powerful new tool for understanding neural mechanisms.
• Facilitates advancements in neurotechnology and broad neuroscience research.
• Advances the modeling of complex neural population dynamics.