Performance of normative and approximate evidence accumulation on the dynamic clicks task

Area of Science:

• Psychophysics
• Decision Neuroscience
• Computational Neuroscience

Background:

• Understanding decision-making in dynamic environments is crucial for psychophysics.
• Mammalian decision-making processes are complex and influenced by changing external factors.

Purpose of the Study:

• To examine ideal and near-ideal observer characteristics in a dynamic decision task.
• To determine how task parameters and design influence observer performance.
• To understand what observer performance reveals about their decision-making strategies.

Main Methods:

• Utilized a dynamic clicks task with two streams of Poisson clicks at varying rates.
• Subjects identified the side with the higher click rate, with unpredictable switches.
• Analyzed performance based on task parameters to differentiate between optimal and near-optimal observers.

Main Results:

• Identified specific task parameter regions where optimal observers maintain constant accuracy, unlike near-optimal ones.
• Demonstrated that approximate normative models require fine-tuning for near-optimal performance.
• Showed that negative log-likelihood and 0/1-loss functions yield different parameter recovery, with 0/1-loss introducing bias amplified by sensory noise.

Conclusions:

• Findings suggest methods to distinguish between closely related normative models.
• Highlights potential pitfalls in experimental design, model fitting, and data interpretation for decision-making studies.
• Provides insights into the characteristics of optimal versus near-optimal decision-making in fluctuating environments.