What Is the Optimal Architecture for Visual Information Routing?

Abstract

Analyzing the design of networks for visual information routing is an underconstrained problem due to insufficient anatomical and physiological data. We propose here optimality criteria for the design of routing networks. For a very general architecture, we derive the number of routing layers and the fanout that minimize the required neural circuitry. The optimal fanout l is independent of network size, while the number k of layers scales logarithmically (with a prefactor below 1), with the number n of visual resolution units to be routed independently. The results are found to agree with data of the primate visual system.

Information Routing, Correspondence Finding, and Object Recognition in the Brain

“…This is not accounted for in our current model, but will be included in future extensions. We have described previously how such a routing over several stages should look like (Wolfrum and von der Malsburg 2007b) and how it can develop ontogenetically (Wolfrum and von der Malsburg 2007a). Finally, the Gallery of our model might correspond to an area like the fusiform face area (FFA), which is specialized for face recognition (Kanwisher andYovel 2006, Tsao et al 2006).…”

“…In Section 3.4, the resulting networks are discussed in functional and physiological terms, before the chapter concludes with Section 3.5. Parts of the material presented in this chapter has been published in (Wolfrum and von der Malsburg 2007b).…”

Rapid Convergence to Feature Layer Correspondences

“…This would require a potentially unrealistic number of axons to converge on one unit. As proposed in Olshausen et al (1993), this problem is likely solved in the brain with the help of intermediate layers to reduce the necessary fan-in and the number of required connections (Wolfrum & von der Malsburg, 2007). Further, whereas in the existing system, the model layer contains just one pattern that is to be compared to the image layer, the object recognition problem has to select the right model from a layer containing dozens of thousands of stored patterns.…”

Self-Organization of Topographic Bilinear Networks for Invariant Recognition

“…These representations, which may reside in inferotemporal cortex (Tanaka, 1996), enable the brain to recognize and examine objects in spite of rapidly changing retinal images. A possible mechanism for the construction of invariant representations is based on variable fiber projections (Hinton, 1981), called dynamic links (von der Malsburg, 1981) or shifter circuits (Anderson & Van Essen, 1987;Wolfrum & von der Malsburg, 2007). The underlying idea of these approaches is to normalize input patterns by applying a transformation to yield a standardized representation.…”

“…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. Nevertheless, for the sake of simplicity, our concrete model lets each global projection pattern be controlled by a single control unit.…”