A Bayesian inference theory of attention: neuroscience and algorithms

AbstractThe past four decades of research in visual neuroscience has generated a large and disparate body of literature onthe role of attention [Itti et al., 2005]. Although several models have been developed to describe specic propertiesof attention, a theoretical framework that explains the computational role of attention and is consistent with all knowneffects is still needed. Recently, several authors have suggested that visual perception can be interpreted as a Bayesianinference process [Rao et al., 2002, Knill and Richards, 1996, Lee and Mumford, 2003]. Within this framework, top-down priors via cortical feedback help disambiguate noisy bottom-up sensory input signals. Building on earlier workby Rao [2005], we show that this Bayesian inference proposal can be extended to explain the role and predict the mainproperties of attention: namely to facilitate the recognition of objects in clutter. Visual recognition proceeds by estimatingthe posterior probabilities for objects and their locations within an image via an exchange of messages between ventraland parietal areas of the visual cortex. Within this framework, spatial attention is used to reduce the uncertainty in featureinformation; feature-basedattentionisusedtoreducetheuncertaintyinlocationinformation. Inconjunction,theyareusedto recognize objects in clutter. Here, we nd that several key attentional phenomena such such as pop-out, multiplicativemodulation and change in contrast response emerge naturally as a property of the network. We explain the idea in threestages. We start with developing a simplied model of attention in the brain identifying the primary areas involved andtheir interconnections. Secondly, we propose a Bayesian network where each node has direct neural correlates within oursimplied biological model. Finally, we elucidate the properties of the resulting model, showing that the predictions areconsistent with physiological and behavioral evidence.