Rapid Convergence to Feature Layer Correspondences

Abstract

We describe a neural network able to rapidly establish correspondence between neural feature layers. Each of the network's two layers consists of interconnected cortical columns, and each column consists of inhibitorily coupled subpopulations of excitatory neurons. The dynamics of the system builds on a dynamic model of a single column, which is consistent with recent experimental findings. The network realizes dynamic links between its layers with the help of specialized columns that evaluate similarities between the activity distributions of local feature cell populations, are subject to a topology constraint, and can gate the transfer of feature information between the neural layers. The system can robustly be applied to natural images, and correspondences are found in time intervals estimated to be smaller than 100 ms in physiological terms.

Dynamic Coordination in Brain and Mind

“…These two mechanisms are not mutually exclusive, and they may indeed be two sides of a coin, signal correlations switching synaptic strengths, synaptic strengths modulating signal correlations (von der Malsburg 1981). Alternately, synapses could be actively controlled by dedicated "control units" (Lücke et al 2008). A third mechanism would require that for each local feature to be signalled, such as a visual edge at a given position and orientation, there were several alternate neurons, each carrying their own distinct connection patterns.…”

Combining Feature- and Correspondence-Based Methods for Visual Object Recognition

“…Their necessity can, however, be argued on the basis of computational theory; this letter is an example. Recently, detailed models have been presented that allow fast correspondence estimation in a neural system (Lücke, Keck, & von der Malsburg, 2008).…”

Self-Organization of Topographic Bilinear Networks for Invariant Recognition

“…Hence, the approach factorizes the representation of input data into the invariant code representing the input independent of the transformation and a code for representing which transformation was needed to achieve this invariance. Invariant object recognition has been modeled on this basis (Olshausen, Anderson, & Van Essen, 1993;Lades et al,1993;Wiskott & von der Malsburg, 1996;Arathorn, 2002;Lücke, Keck, & von der Malsburg, 2008;Wolfrum, Wolff, Lücke, & von der Malsburg, 2008). Dynamic links may be controlled by temporal correlations of neuronal activities in a rapid reversible version of Hebbian learning, and face recognition has been modeled in this way (Wiskott & von der Malsburg, 1996).…”

“…If control units responsible for alternate transformations compete in a winner-take-all fashion, the stage is set for very rapid transformation detection and subsequent transformationspecific signal routing. This leads to the compensation of variations of input patterns and generates an invariant representation that can be exploited for object recognition (Olshausen et al, 1993;Arathorn, 2002;Lücke et al, 2008;Wolfrum et al, 2008).…”

“…As proposed in Anderson and Van Essen (1987) and Zhu and von der Malsburg (2004), the RPFs of control units should be intermediate between controlling single links, as in Lücke et al (2008) and Wolfrum et al (2008), and the total set of connections involved in a transformation from the input to the whole invariant output window. The latter, though most efficient (in that a single active control unit could project a whole figure into an invariant output window in inferotemporal cortex) is unrealistic for several reasons: due to the limited spatial range of neurites; because a whole projection has to traverse several cortical areas (e.g., V1-V2-V4-IT), as modeled in Anderson and Van Essen (1987) and Wolfrum and von der Malsburg (2007); and since the number of required control units would be too large to cover the space of all possible projection patterns.…”