Rapid Correspondence Finding in Networks of Cortical Columns

Rapid Convergence to Feature Layer Correspondences

“…First, the feature units with the smallest inputs are deactivated, and finally only subsets of units with strong input remain active. During map formation, however, many feature units are always active (compare Lücke & von der Malsburg, 2006). Only for very high levels of inhibition, and after the map has formed, just one unit per feature column remains active.…”

“…For simplicity, the feature layers thus solely communicate via their systems of control columns in the dynamics studied here. In a more general system, input to feature columns (see equation 3.1) can, however, also constitute of mixtures of feature vector input and feature column input from the other layer (compare Lücke & von der Malsburg, 2006).…”

“…In this letter, we study the neural dynamics of link selection, for which the direct connections between feature columns are not required. For simplicity, the inputs in equation 3.1 will therefore convey information from the (input or model) image alone (see Lücke & von der Malsburg, 2006, for a system with about an equal mixture of image input and input from the other layer). Details of the image features will be given in section 5.…”

“…The matrices W M i,I j and W I i,M j represent the connectivity from input layer I to model layer M and from M to I, respectively. For one-dimensional layers, the matrices W are two-dimensional, and their visualization allows a direct interpretation of the interlayer connectivity (compare Zhu & von der Malsburg, 2004;Lücke & von der Malsburg, 2006). 3 For two-dimensional input, W M i,I j and W I i,M j are four-dimensional matrices, and although their entries can in principle be visualized, the connectivities they represent are difficult to grasp intuitively.…”

“…Similar to diagonal connections, a connectivity that involves connections between approximately parallel fibers has not produced significant functional improvements for the sets of natural images considered here (see below). In a system with one-dimensional layers as described in Lücke and von der Malsburg (2006), approximately parallel links are also interacting. There, feature vectors were artificial and independent from node to node.…”