Attentional Processes in Correspondence-Based Object Recognition

We show here that a model for invariant object recognition described earlier [1] naturally displays effects of space based and object based attention. The network consists of three functional layers: an input layer for image representation, an intermedi-ate assembly layer for recurrent information integra-tion, and a gallery layer for memory storage (see figure). Each layer consists of cortical columns as functional building blocks that are modeled in ac-cordance with recent experimental findings. Con-nections between input and assembly layer are mod-ulated by control units on a fast time scale to allow detection of objects in the input image and position invariant mapping to the normalized assembly layer. Recurrent information exchange between assembly and gallery layer leads to recognition. Tests on stan-dard benchmark databases show competitive perfor-mance for face recognition. Here, we prepare the system to recognize faces of 1000 different identities invariantly of position. It is then confronted with a large test image containing several of those faces (under different expression). We show how by pre-activating a subset of the control units in a top-downmanner (supposedly from prefrontal cortex), attention of the system can be directed to a certain part of visual space, resulting in preferred processing and recognition of objects at that position. This bias can be neutralized by very salient stimuli at other positions of the visual field. Similarly, top-down activation of specific subsets of the memory domain (e.g. female faces or a single identity) results in object priming: Among several faces in the visual field, the system will preferentially detect and subsequently recognize those faces indicated by the subset. In accord with [2], selection of a location in the input domain is not achieved by increased activity at the input level but by selective routing to higher domains, resulting in effectively shifting receptive fields of neurons at higher stages. In distinction to previous work (e.g. [3]), our system keeps track of spatial relationships between features and is therefore able to distinguish between objects of similar qualitative structure (i.e. containing many similar features)