A probabilistic strategy for understanding action selection.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Decision Neuroscience

Background:

• Decision formation occurs in brain regions processing sensory-motor transformations.
• The mechanism of reading out decisions from neural populations to generate action choices is not fully understood.

Purpose of the Study:

• To investigate the computational principles of population coding for action selection.
• To compare the predictive accuracy of different models for action choice.

Main Methods:

• Simultaneous recording of four superior colliculus neurons in monkeys during a target selection task.
• Implementation and comparison of three models: population vector average (PVA)/optimal linear estimator (OLE), winner-takes-all (WTA), and maximum a posteriori (MAP) Bayesian model.

Main Results:

• The MAP Bayesian model demonstrated superior prediction accuracy (81.88%) compared to WTA (71.11%) and PVA/OLE (55.71%/69.47%).
• Incorporating nonuniform priors improved MAP model accuracy by 2.88%.
• MAP estimates evolved dynamically, peaking at saccade initiation and scaling with performance.

Conclusions:

• Probabilistic frameworks, specifically the MAP model, provide a robust understanding of action selection from neural population activity.
• The study highlights the importance of dynamic and probabilistic computations in decision-making and motor control.