Interrogating theoretical models of neural computation with emergent property inference

Area of Science:

• Theoretical Neuroscience
• Computational Neuroscience
• Systems Neuroscience

Background:

• Neural circuit models are crucial for understanding neural computation and linking theory to experimental observations.
• Model parameter identification is a key challenge in theoretical neuroscience, often involving solving complex inverse problems.
• Existing methods may struggle with the high dimensionality and specific computational properties required for accurate neural modeling.

Purpose of the Study:

• To present a novel technique, emergent property inference (EPI), for solving the inverse problem in theoretical neuroscience.
• To leverage deep neural networks within a probabilistic modeling framework for parameter inference.
• To demonstrate EPI's applicability and advantages over existing methods in neuroscience.

Main Methods:

• Developed emergent property inference (EPI), a novel technique integrating deep neural networks and probabilistic modeling.
• Applied EPI to learn parameter distributions that capture specific computational properties of neural circuit models.
• Utilized a motivational example of parameter inference in the stomatogastric ganglion to introduce and validate the methodology.

Main Results:

• EPI successfully infers parameters for neural circuit models, demonstrating precise control over inferred parameter behavior.
• The method exhibits superior scalability in parameter dimension compared to alternative techniques.
• Novel theoretical insights were gained in models of the primary visual cortex and superior colliculus through EPI analysis.

Conclusions:

• Emergent property inference (EPI) offers a powerful, scalable approach to neural circuit model parameter identification.
• The integration of deep learning with probabilistic modeling opens new avenues for theoretical neuroscience research.
• EPI facilitates the discovery of complex parametric structures and advances our understanding of neural computation.