Testing and improving the robustness of amortized bayesian inference for cognitive models.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Statistical Modeling

Background:

• Parameter estimation in cognitive models is susceptible to contaminant observations.
• Robustness is crucial for reliable model parameter estimation, especially with noisy data.

Purpose of the Study:

• To improve the robustness of parameter estimation in cognitive models using amortized Bayesian inference.
• To develop and evaluate a data augmentation method for training robust estimators.

Main Methods:

• Systematic analyses were conducted on a toy normal distribution example and the drift diffusion model.
• A data augmentation approach incorporating a contamination distribution (Cauchy) during training was proposed.
• Robust estimators were evaluated for accuracy and efficiency loss compared to a standard estimator.

Main Results:

• Introducing Cauchy contaminants during training significantly enhanced the robustness of the neural density estimator.
• Robust estimators demonstrated bounded sensitivity functions and a substantially higher breakdown point.
• The proposed method showed practical implementation with minimal accuracy and efficiency loss.

Conclusions:

• Amortized Bayesian inference combined with data augmentation offers a practical solution for robust parameter estimation.
• The method is broadly applicable in fields facing challenges with outlier detection and removal.
• This approach enhances the reliability of cognitive model parameter estimation in the presence of contaminants.