Related Work - Hierarchical Neural Networks for Image Interpretation - page 63

Top-down attention

V1 layer 2/3 activity

Visual input

(b)

(c)
Fig. 3.21. Grouping and attention in the LAMINART model proposed by Grossberg [84]:
(a) architecture of the system (feed-forward, feedback, and horizontal interactions within
and between LGN, V1, and V2; open symbols indicate excitation; closed symbols represent
inhibition); (b) top-down attention and bottom-up stimuli interact (the attention spreads along
the illusory contour); (c) perceptual grouping (left: visual input; right: V1 response; top:
vertical grouping; bottom: horizontal grouping) (images from [183, 85]).

(a)

is not clear, how to extend the model to represent more complex features in higher
visual areas.
LAMINART Model. Grossberg [84] proposed a model for the laminar circuits in
V1 and V2 that he called LAMINART. This model accounts for perceptual effects,
such as grouping, orientation contrast, and attention. It is based on lateral and verti-
cal feedback connections that have been found in the cortex. The architecture of the
model is sketched in Figure 3.21(a).
Three visual areas are arranged in a hierarchical fashion. LGN is connected to
V1 via vertical feedback loops. The same connectivity pattern exists between V1
and V2. On-center/off-surround type interactions between adjacent layers imple-
ment an ART-like resonance [40] between features of different complexity. Corre-
sponding features strengthen each other, while incompatible features are inhibited.
Within an area, horizontal connections facilitate perceptual grouping of collinear
cells that have the same orientation. The range of these connections is larger in V2
than in V1. Figure 3.21(c) shows the grouping results of two stimuli. While in the