engleski

A Multi-Scale Neural Architecture for Incremental Grouping

Houtkamp and Roelfsema (2011) suggested that perceptual grouping consist of two processes: 1) base grouping that rely on a fast extraction of image features occurring in parallel across the visual scene and 2) incremental grouping that rely on slow, serial labeling of image elements that belong to perceptual group. At the neural level, the incremental grouping involves spreading of enhanced firing rate across neurons encoding the segments of the same contour (Roelfsema, 2006). Here, we developed a multi- scale neural architecture to explain how early visual cortex implements incremental grouping and object-based attention. We employed three different spatial scales arranged in an image pyramid with feedforward and feedback connections between scales. At each scale, a set of oriented filters is used to simulate the properties of cortical cells with oriented receptive fields. Within each scale, there are recurrent excitatory connections among nodes with collinear receptive fields. Furthermore, each scale implements iso- orientation and cross-orientation inhibition among nearby spatial locations. Computer simulations showed that the enhanced neural activity spreads along the contour segments with variable speed depending on the distance between attended and unattended contour and on their curvature (Jolicoeur et al., 1986 ; 1991). Also, we showed how activity spreading occur across X- and T-junctions without spill-over to unattended contour. Finally, we showed that the model helps explain Gestalt principle of good continuation. In conclusion, we demonstrated that the proposed neural architecture can serve as a cognitive blackboard to implement visual routines as proposed by Roelfsema and de Lange (2016).

ncremental grouping, object-based attention, neural networks

nije evidentirano