Rüdiger Kupper, Marc-Oliver Gewaltig, Andreas Knoblauch, Ursula Körner, Edgar Körner, "From neurons to cortex: a multi-level approach to understanding the brain", Neurocomputing Research Trends, Nova Science Publishers, Inc., 2007.

At the Honda Research Institute, we aim to understand the operating principles of the brain. We think that the cortex is composed of elementary building blocks, the columns, that apply one generic algorithm to varying sensory data. The brain is thus not a collection of highly specialized neural circuits, providing tailored solutions to individual problems, but it uses the same set of powerful processing strategies over and over again. Based on neurobiological knowledge about the primate cortex we substantiate this idea in a model of the visual system, on several levels of detail. At the single neuron level, the visual system can immediately profit from a spike-latency code to rapidly segment and recognize scenes. A wave of spikes traveling through a cascade of feature detectors rapidly activates a high-level hypothesis about stimulus content, so that an appropriate reaction (e.g. escape) is possible. At the level of neural circuits, we simulate the signal flow in columns, across the six cortical layers and between several cortical areas. We show how the columnar subsystems interact to predict and recognize stimuli in terms of acquired knowledge. Columnar communication integrates top-down and bottom-up signals to describe the stimulus consistently across all cortical areas. It iterates this process to refine the description, causing oscillations in neural activity. Internal descriptions of entirely new stimuli can be constructed from old ones, and entrained with the help of the hippocampal formation. At the system level, we implement a large-scale model of main visual cortical areas, parts of the hippocampal formation, and sub-cortical structures. Cortical columns both predict future stimuli and vote for motor actions to confirm them. Here, they control saccades to learn and recognize objects from a sequence of partial views, based on step-by-step prediction and refinement of an object hypothesis in time. In our models the columns link the neural level to the system level. They help us to understand how groups of nerve cells ultimately create the macroscopic function of the brain.