Multi-step planning of eye movements in visual search

Area of Science:

• Cognitive Neuroscience
• Computational Vision
• Human Factors

Background:

• Gaze selection is vital for survival and has been modeled computationally.
• Previous models focused on immediate maximal criteria (e.g., reward, uncertainty reduction) or single-step planning.
• Tasks requiring extended action sequences necessitate planning beyond a single gaze shift.

Purpose of the Study:

• To investigate if the human visual system plans eye movement sequences beyond the next single gaze shift.
• To provide quantitative evidence for optimal probabilistic planning in human gaze selection.
• To compare human behavior against myopic (single-step) and planning observer models.

Main Methods:

• Employed a series of computational models to analyze human gaze selection in visual search tasks.
• Quantitatively compared subject behavior to predictions from a planning observer model versus a myopic observer model.
• Varied visual stimuli and available search time to assess planning flexibility.

Main Results:

• Human gaze behavior was better explained by a planning observer model than a myopic, greedy observer model.
• The location of the first eye fixation varied based on stimulus and time constraints, as predicted by the planning model.
• Demonstrated quantitative agreement between human gaze selection and optimal planning under uncertainty.

Conclusions:

• Provides the first quantitative evidence that human gaze selection involves planning beyond a single eye movement.
• Suggests the human visual system employs probabilistic planning strategies for optimal eye movement sequences.
• Highlights the importance of considering multi-step planning in models of visual search and attention.