Neural Abstraction Pyramid Architecture - Hierarchical Neural Networks for Image Interpretation - page 65

4. Neural Abstraction Pyramid Architecture
The last two chapters reviewed what is known about object recognition in the human
brain and how the concepts of hierarchy and recurrence have been applied to image
processing. Now it is time to put both together.
In this chapter, an architecture for image interpretation is defined that will be
used for the remainder of this thesis. I will refer to this architecture as the Neu-
ral Abstraction Pyramid. The Neural Abstraction Pyramid is a neurobiologically
inspired hierarchical neural network with local recurrent connectivity. Images are
represented at multiple levels of abstraction. Local connections form horizontal and
vertical feedback loops between simple processing elements. This allows to resolve
ambiguities by the flexible use of partial interpretation results as context.
4.1 Overview
Before going to the details, this section gives an overview of the proposed architec-
ture. It covers the hierarchical network structure, the use of distributed representa-
tions, local recurrent connectivity, and the idea of iterative refinement.
4.1.1 Hierarchical Network Structure
As the name implies, the Neural Abstraction Pyramid has a hierarchical net-
work structure. It is sketched in Figure 4.1. The network consists of several two-
dimensional layers that represent images at different degrees of abstraction. Each
layer is composed of multiple feature arrays that contain discrete cells, called feature
cells. When going up the hierarchy, the number of feature arrays per layer increases,
while the spatial resolution decreases.
Unlike most neural networks that have no spatial organization within the layers,
layers in the Neural Abstraction Pyramid have a two-dimensional organization that
corresponds to the 2D nature of images. This is motivated by the observation that
correlations between image locations that are close together are higher than correla-
tions between locations that are far-apart. This simple fact is prewired in the network
structure in the same way as it is prewired in the retinotopic organization of cortical
areas in the human visual system (compare to Chapter 2).
The hierarchical network architecture resembles the hierarchy of areas in the
ventral visual pathway (see Section 2.1). The idea of a stack of 2D layers with