The different stages of visual recognition need different attentional binding strategies

A Hierarchical Generative Model of Recurrent Object-Based Attention in the Visual Cortex

“…We found classification and reconstruction performance to be reasonable after just the initial FF pass on non-cluttered data sets. For cluttered images however, subsequent recurrent processing was important to achieve better object specific representations, in line with what is suggested for the cortex [8,5].…”

“…Some theories pose that the aim of attentional processing is to form representations that are specific to one object at a time, especially in higher cortical areas [10,5]. We thus trained the models on individual objects only (shapes or digits), but then tested them on the various cluttered data sets to see whether information about individual objects is retrieved in the highest hidden layer even when scenes are complex in ways not seen in training.…”

“…When the model is run (performing Gibbs sampling on the joint probability), its state performs a random walk in the energy landscape along basins of attraction, which embody meaningful representations obtained during training. Because the latter are specific to individual objects by construction, the model assumes (stochastic) attractor states representing the objects being attended to [9,5], as shown below.…”

“…In that sense, this work can be understood as modeling in particular covert attention. It relates to approaches such as Selective Tuning [5] and others (e.g. [6,7]), but is unique in capturing facets of attention in a framework implementing aforementioned general properties.…”

“…We qualitatively elucidate on the following aspects of theories of attentional processing in the cortex: First, the notion of a fast feed-forward (FF) sweep followed by subsequent recurrent processing, the latter being essential for perceiving objects when scenes are cluttered [8]; second, that, in directing attention to an individual object in a scene, an attractor state is assumed which binds together and emphasizes aspects of that object represented throughout the cortical hierarchy, suppressing representations of competing objects [9,5]; third, the hypothesis that scene representations in the cortex are inherently such that higher stages represent primarily one object at a time, unlike lower stages such as V1 where the whole image is encoded in terms of low-level features [10].…”