Uncovering Dynamical Equations of Stochastic Decision Models Using Data-Driven SINDy Algorithm

Area of Science:

• Cognitive Neuroscience
• Computational Neuroscience
• Dynamical Systems

Background:

• Perceptual decision making relies on accumulating sensory evidence over time.
• Sequential sampling models describe this process, with decision variables tracked by neural activity.
• Current computational methods for analyzing decision dynamics are limited.

Purpose of the Study:

• To apply sparse identification of nonlinear dynamics (SINDy) to uncover deterministic components of stochastic decision models.
• To evaluate SINDy's effectiveness in estimating model parameters and predicting behavior from simulated neural data.
• To explore SINDy's utility for analyzing perceptual decision-making dynamics.

Main Methods:

• Utilized sparse identification of nonlinear dynamics (SINDy), a data-driven approach.
• Applied SINDy to simulated decision variable activities from reaction time tasks.
• Investigated multi-trial, trial-averaging, and single-trial SINDy approaches, assuming known noise coefficients.

Main Results:

• SINDy successfully estimated deterministic terms in dynamical equations, choice accuracy, and decision time across various signal-to-noise ratios.
• The multi-trial SINDy approach yielded the best performance.
• Single-trial SINDy showed potential for real-time modeling.

Conclusions:

• SINDy offers a powerful data-driven method for elucidating the dynamics of perceptual decision making.
• The findings provide alternative approaches for analyzing first-passage time problems using SINDy.
• This work advances computational methods for understanding neural mechanisms of choice.