Salience maps in parietal cortex: imaging and computational modeling.

Models of spatial attention are often based on the concept of a salience map. In computational cognitive neuroscience, such maps are implemented as a collection of nodes with self-excitation and lateral inhibition between all nodes (competitive interaction map). Here, we test some critical predictions of this idea. We argued that task demands, more precisely the level of attention required, can top-down modulate the level of lateral inhibition in a salience map, and thus induce different activation functions. We first show that a model with a high lateral inhibition parameter generates a monotonous activation curve as a function of set size similar to that typically observed in the literature (e.g. Todd and Marois, 2004). Next, we show that a competitive interaction map with medium lateral inhibition leads to a Lambda-shaped activation curve when set sizes increase. This prediction is confirmed in an fMRI experiment with medium attention demands where a similar Lambda-shaped activation curve is found in a posterior superior parietal area that was proposed to house a salience map (Todd and Marois, 2004). Finally, we show that a qualitatively different V-shaped activation curve is predicted with a very low inhibition parameter. An fMRI experiment with low attentional demands revealed this V-shaped activation curve in the same region. These findings provide critical support for the existence of a salience map based on competitive interactions in posterior superior parietal cortex, and suggest that its parameters (in particular, lateral inhibition) can be modulated in a top down manner dependent on task demands.