Model discovery to link neural activity to behavioral tasks.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Systems Neuroscience

Background:

• Brains evolve through natural selection, not engineering, making direct modeling of neural activity challenging.
• Existing models may not fully capture the complex relationship between neural activity and experimental conditions.

Purpose of the Study:

• To develop an accessible framework, Model Identification of Neural Encoding (MINE), for discovering and characterizing models of neural encoding.
• To interpret these models using Taylor decomposition to understand how task features map to neural activity.

Main Methods:

• Utilized convolutional neural networks (CNNs) within the MINE framework to build predictive models.
• Applied Taylor decomposition techniques to interpret the complex CNN models.
• Tested MINE on published cortical data and zebrafish thermoregulatory circuit experiments.

Main Results:

• MINE successfully characterized neurons by receptive field and computational complexity, revealing anatomical segregation.
• Identified a novel class of neurons integrating thermosensory and behavioral information.
• Demonstrated MINE's ability to uncover neuronal functions missed by traditional methods.

Conclusions:

• MINE provides a powerful and interpretable approach to understanding neural encoding.
• The framework advances the characterization of neuronal function and discovery of new neural circuit principles.
• MINE offers a significant improvement over traditional clustering and regression for analyzing complex neural datasets.