engleski

A Model of Multiple Object-Tracking

Pylyshyn and Storm (1988) showed that visual system individuates simple objects and track their movement in visual field. Such ability has limited capacity because it is possible to track only four to five objects simultaneously. Tracking does not depend on the serial scanning strategy but involves parallel mechanism which assigns pointers to objects in the spatial representation. These pointers move along with objects and they are named fingers of instantiation (FINST). Existence of FINST implied that focus of spatial attention is divided into several independent channels (Cavanagh & Alvarez, 2005). On the other hand, Yantis (1992) argued that tracking is performed by forming perceptual groups of moving objects. When the trajectories of the objects allow perceptual grouping, tracking performance increased. We proposed a recurrent neural network for visual working memory which is able to track changes in the location of simple input patterns. The model is based on self-excitation and two types of inhibition: recurrent (or lateral) and dendritic inhibition. Dendritic inhibition protects nodes in the network from lateral competitive influences. Due to the dendritic inhibition, node in the network will receive lateral inhibition only from other nodes with higher activity level. Asymmetric competition induces ordering of neural activity, in a way that different objects are labeled with different activity amplitudes. Separate active nodes might be considered as an implementation of multi-focal attention (Cavanagh & Alvarez, 2005). Furthermore, input signals are modulated by the network activity in order to keep segregated activity labels for different objects when they change locations. Computer simulations showed that the neural activity in the model has an ability to track limited number of independently moving objects. Capacity limitation arises from the limited dynamic range of model neurons. Ability to track moving objects arises from the network sensitivity to new input signals. When object appears at new location, neural activity at that location overrides activity corresponding with previous object location due to the asymmetric inhibition. Tracking is compromised if two objects are close to each other because neural activity corresponding to them merges and becomes indistinguishable for further tracking. However, as shown by Viswanathan and Mingolla (2002), merge between object labels does not occur if objects are perceived as separated in depth either by occlusion or by stereo disparity. The model account for this finding with the assumption that different depth planes are represented with different network layers which compete among each other. Self-excitation enables tracking of objects which disappear and reappear behind visible or invisible occluded surfaces. When moving objects form a perceptual group, we assume that they received the same activity label in the neural network. Due to the dendritic inhibition, nodes with the same activity level will not compete and the network will be able to override capacity limitation imposed by lateral inhibition. Therefore, the model provides unified account of competitive and cooperative influences on multi-object tracking.

Attention; Object-Tracking; Capacity Limitation; Visual Working Memory; Neural Model

nije evidentirano