UCL Logo

Angela Yu

Department of Cognitive Science, UCSD, USA


Tuesday 1 June 2010



Seminar Room B10 (Basement)

Alexandra House, 17 Queen Square, London, WC1N 3AR


Optimal decision-making in an active visual search task: effects of spatial statistics on motor planning and sensory processing


Humans and animals are often faced with sensory processing and action planning under conditions of uncertainty, a daunting task aided by the discovery and exploitation of statistical regularities in the environment. We examine the learning and utilization of such statistical regularities in an active visual search task, in which subjects make sequential fixation decisions to locate a hard-to-identify motion target embedded among distractors. We impose different spatial biases in target location across different blocks of trials, and examine how subjects' search strategy adapts to such spatial biases. In theory, for subjects to internalize and exploit statistical regularities in target location, they must integrate and extract the relevant information across trials. Moreover, within each trial, there needs to be an integral and dynamic interaction between sensory processing and action planning: noisy visual inputs partially inform subsequent saccade planning, while fixation choices impact the nature and fidelity of future sensory inputs. We use Bayesian statistical theory and stochastic control theory to formulate an "ideal searcher" model of visual search strategy that maximizes experimentally defined rewards (i.e. target detection accuracy) and minimizes costs (i.e. search time, fixation location switch cost), by (A) extracting relevant spatial statistics across trials, and (B) sequentially and adaptively enact sensory-motor decisions within trials. This ideal searcher model makes specific predictions on fixation location and duration, as well as accuracy and reaction time of target detection, which we compare to human behavior. We show that subjects indeed internalize statistical regularities in the environment, and exploit such regularities in both sensory processing and action planning. For example, on the sensory side, they are biased to perceive the target as being located in the most probable location, whether the target is truly there or not. They also need less sensory processing time to confirm target presence in the expected location, and more time to overcome prior bias when the target is not there, both when compared to less likely locations for the target. On the motor planning side, their first fixation on each trial is strongly biased toward the most likely target location, while subsequent fixations are biased toward less frequent target locations in a graded manner. Combined with our modeling results, these data indicate that subjects are near-optimal in learning about statistical regularities at multiple time-scales, and utilizing such statistical prior knowledge to mediate sensory-motor processing in active visual search.