Flexible and efficient simulation-based inference for models of decision-making.

Area of Science:

• Cognitive Neuroscience
• Computational Modeling
• Bayesian Inference

Background:

• Parameter inference for computational models is crucial in cognitive neuroscience.
• Traditional Bayesian methods struggle with models lacking efficient likelihood computation.
• Simulation-based inference (SBI) bypasses likelihood evaluation but can be simulation-intensive.

Purpose of the Study:

• To develop a more simulation-efficient method for applying SBI to cognitive models.
• To enable accurate Bayesian parameter inference for complex decision-making models.
• To facilitate faster model design and scientific discovery in cognitive neuroscience.

Main Methods:

• Introduced Mixed Neural Likelihood Estimation (MNLE), a novel SBI approach.
• MNLE trains neural density estimators on model simulations to emulate the simulator.
• The method captures both continuous (e.g., reaction times) and discrete (choices) data.

Main Results:

• MNLE achieves comparable likelihood accuracy to Likelihood Approximation Networks (LANs) using six orders of magnitude fewer simulations.
• MNLE demonstrates significantly higher accuracy than LANs when trained with equivalent simulation budgets.
• Validated MNLE on two variants of the drift-diffusion model.

Conclusions:

• MNLE offers a substantially more simulation-efficient alternative for SBI in cognitive neuroscience.
• This method allows for Bayesian parameter inference using standard approximate methods like Markov Chain Monte Carlo sampling.
• MNLE empowers researchers to efficiently analyze custom decision-making models, accelerating scientific exploration.