One dimensional approximations of neuronal dynamics reveal computational strategy

Area of Science:

• Neuroscience
• Computational Neuroscience
• Artificial Intelligence

Background:

• The link between neural activity and the computations it performs is not well understood.
• Existing models often lack simplicity and interpretability.

Purpose of the Study:

• To develop a novel methodology for modeling neuronal activity.
• To establish a link between neural activity patterns and computational strategies.
• To compare computational strategies across biological and artificial systems.

Main Methods:

• Developed a method to condense neuronal activity into interpretable 1-dimensional trajectory models.
• Validated the model across diverse systems, including C. elegans neurons and human fMRI.
• Analyzed the structure of interconnected trajectories (scaffolds) to infer computational strategies.

Main Results:

• The trajectory models accurately predict future neuronal activity and human decisions.
• The scaffold structure reflects the system's computational strategy.
• Identified conditions under which artificial agents adopt primate-like computational strategies.
• The extracted computational strategies predict errors on novel stimuli.

Conclusions:

• The developed methodology provides a powerful tool for studying the computation-activity relationship.
• This approach enables cross-system comparisons of computational strategies.
• Offers insights into how artificial systems learn and execute strategies.