Cost-sensitive Bayesian control policy in human active sensing.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Systems Neuroscience

Background:

• Active sensing, utilizing self-motion for focused sensory input, is crucial but not well understood.
• Existing models often optimize abstract objectives like information gain, not explicit behavioral costs.

Purpose of the Study:

• To present a novel Bayesian model, C-DAC (Context-Dependent Active Controller), for active sensing.
• To compare C-DAC's performance against previous models in explaining human visual search dynamics.
• To explore efficient approximations of C-DAC for biological plausibility and engineering applications.

Main Methods:

• Collected behavioral data from a visual search experiment.
• Developed a Bayesian inference and control framework (C-DAC) minimizing temporal delay, response error, and sensor repositioning costs.
• Compared C-DAC with existing algorithms like Infomax and greedy MAP.

Main Results:

• C-DAC demonstrates superior performance in capturing human visual search dynamics compared to prior models.
• The model effectively accounts for "confirmation bias" in how humans use prior spatial knowledge to enhance search.
• Efficient approximations of C-DAC show potential for explaining neural computations and engineering solutions.

Conclusions:

• C-DAC provides a more accurate framework for understanding active sensing and human visual search.
• The model highlights the importance of minimizing explicit behavioral costs in active sensing.
• Investigated approximations offer insights into neural mechanisms and practical applications of active sensing.