Non-Parametric Bayesian Modelling of Human Visual Attention

 

Matthew Rounds


Organisms attempting to make sense of the world are faced with a difficult task; they must infer the structure and content of an environment that is constantly changing, and they must do so in the face of limited processing power, and with limited synchronic access to said environment (Duncan 1980, Huang et al 2007).


One of the primary tools available to an organism is its control over which part of its environment it accesses and, internally, which features of the resultant sensorium it considers most relevant to the task of inference. Both of these can be considered types of 'attention', which for the moment we will simply define as 'selection for action'. The two are tightly bound, and under some accounts, a result of the same mechanism (Clark, 2013). 


The role of attention in human cognition has long been a source of interest: both in philosophy (e.g. Dainton 2004, Tye 2003, Watzl 2012), and in psychology (e.g. Huang et al 2007, Franconeri 2005). Attention has been argued to play a significant role in structuring conscious experience, object binding, and even introspection. In neuroscience, it has been suggested to play an important role in the process by which the brain performs inference on the causes of its sensory input (e.g. Friston et al 2012, Hohwy 2012).


The approach of this project is to consider the interplay between covert and overt attention in the context of conjunction search - where performance decreases if distractors share features with the target object (Shen et al 2003), and inattentional blindness - where an unexpected object is more likely to be spotted if its feature overlap those which are task relevant (Jesen et al 2011).


We propose a novel model of these results as both a product of the relationship between a top-down approximately bayesian model of the environment, which favours parsimony in its explanations; and the optimally informative policy of samplings it subsequently chooses. Further work might include modelling the relationship between expertise and inattentional blindness. 

Supervisors: Frank Keller & Chris Lucas

Previous: Music Decomposition