Bayesian and Discriminative Models for Active Visual Perception across Saccades.

Area of Science:

• Neuroscience
• Cognitive Science
• Vision Science

Background:

• The brain must maintain a stable visual percept despite sensory disruptions caused by self-generated actions, such as saccadic eye movements.
• Prior expectations influence visual stability, suggesting a potential role for Bayesian inference in active perception.

Purpose of the Study:

• To test whether active perception, specifically visual stability during saccades, follows Bayesian principles.
• To investigate how sensory uncertainty affects the use of prior expectations in human and macaque perception.
• To determine if different types of uncertainty (external image noise vs. internal motor noise) lead to different perceptual strategies.

Main Methods:

• Humans and rhesus macaques performed tasks reporting visual motion during saccades.
• Prior expectations and sensory uncertainty (image noise and motor noise) were experimentally manipulated.
• Psychophysical data were compared against predictions from Bayesian ideal observer models and a discriminative learning model.

Main Results:

• Humans exhibited Bayesian behavior for continuous visual judgments but anti-Bayesian behavior for categorical judgments, using priors less with increased uncertainty.
• Macaques showed anti-Bayesian behavior for externally induced sensory uncertainty (image noise) but Bayesian behavior for internally driven uncertainty (motor noise).
• A discriminative learning model successfully explained the observed anti-Bayesian effects.

Conclusions:

• Active vision employs both Bayesian and discriminative models, with the choice depending on task demands (continuous vs. categorical) and the source of sensory uncertainty.
• The findings highlight the flexibility of perceptual systems and offer insights into the neural organization underlying active perception by comparing Bayesian and non-Bayesian models.