Herbert J. Reitboeck

Coherent oscillations: A mechanism of feature linking in the visual cortex?

Primary visual coding can be characterized by the receptive field (RF) properties of single neurons. Subject of this paper is our search for a global,second coding step beyond the RF-concept that links related features in a visual scene. In recent models of visual coding, oscillatory activities have been proposed to constitute such linking signals. We tested the neurophysiological relevance of this hypothesis for the visual system. Single and multiple spikes as well as local field potentials were recorded simultaneously from several locations in the primary visual cortex (A17 and A18) using 7 or 19 individually advanceable fibermicroelectrodes (250 or 330 μm apart).Stimulusevoked (SE)-resonances of 35–85 Hz were found in these three types of signals throughout the visual cortex when the primary coding channels were activated by their specific stimuli. Stimulus position, orientation, movement direction and velocity, ocularity and stationary flicker caused specific SE-resonances.Coherent SE-resonances were found at distant cortical positions when at least one of the primary coding properties was similar. Coherence was found1) within a vertical cortex column,2) between neighbouring hypercolumns, and3) between two different cortical areas. We assume that the coherence of SE-resonances is mediated by recurrent excitatory intra- and inter-areal connections via phase locking between assemblies that represent the linking features of the actual visual scene. Visually related activities are, thus, transiently labelled by a temporal code that signalizes their momentary association.

A model for size- and rotation-invariant pattern processing in the visual system

The mapping of retinal space onto the striate cortex of some mammals can be approximated by a log-polar function. It has been proposed that this mapping is of functional importance for scale-and rotation-invariant pattern recognition in the visual system. An exact log-polar transform converts centered scaling and rotation into translations. A subsequent translation-invariant transform, such as the absolute value of the Fourier transform, thus generates overall size-and rotation-invariance. In our model, the translation-invariance is realized via the R-transform. This transform can be executed by simple neural networks, and it does not require the complex computations of the Fourier transform, used in Mellin-transform size-invariance models. The logarithmic space distortion and differentiation in the first processing stage of the model is realized via “Mexican hat” filters whose diameter increases linearly with eccentricity, similar to the characteristics of the receptive fields of retinal ganglion cells. Except for some special cases, the model can explain object recognition independent of size, orientation and position. Some general problems of Mellin-type size-invariance models-that also apply to our model-are discussed.

A neural network for scene segmentation by temporal coding

Abstract We propose a neural network for object definition and scene segmentation via temporal signal correlations. The network consists of model neurons with a dynamic threshold and with two functionally different types of synapses. Activities of model neurons responding to a given object in the input image are synchronized, and desynchronized with respect to the activities of neural assemblies that respond to other objects. The present network can separate up to eight identical objects in a visual scene. The simulation results support the hypothesis that synchronous assembly activity in the visual system serves as a functional principle for feature linking, object definition, and scene segmentation.

Contour Segmentation with Recurrent Neural Networks of Pulse-Coding Neurons

The performance of technical and biological vision systems crucially relies on powerful processing capabilities. Robust object recognition must be based on representations of segmented object candidates which are kept stable and sparse despite the highly variable nature of the environment. Here, we propose a network of pulse-coding neurons based on biological principles which establishs such representations using contour information. The system solves the task of grouping and figureground segregation by creating flexible temporal correlations among contour extracting units. In contrast to similar previous approaches, we explicitly address the problem of processing grey value images. In our multi-layer architecture, the extracted contour features are edges, line endings and vertices which interact by introducing facilatory and inhibitory couplings among feature extracting neurons. As the result of the network dynamics, individual mutually occluding objects become defined by temporally correlated activity on contour representations.

Size and position invariant visual representation supports retinotopic maps via selective backward paths: A dynamic second order neural network model for a possible functional role of recurrent connections in the visual cortex

Abstract A dynamic multi-layered neural network performing scale and position invariant pattern recognition has been simulated. The model consists of a set of retinotopically organized layers that label objects via temporal signal correlations (synchronization coding). The successive set of layers computes scale and position invariance and functions as a dynamic associative memory in which objects can invariantly be stored and retrieved. Model neurons are arranged in six layers and coupled in forward direction via feeding connections and in lateral and backward directions via linking connections. Feeding connections manage the function of feature extraction and pattern recognition while linking connections support synchronizations among mutually coupled neurons. The most important property of the associative memory is its ability to support the retinotopic representation of a visual object via selective backward paths. For this, a special backward network has been included that is selectively activated by the forward pathway performing the invariance operation from the retinotopic layers to the associative memory. If a certain object in the input pattern is recognized by the associative memory, the synchronized signals of the objects memory neurons are routed back selectively to that position of the retinotopic network where the object is actually represented. Due to this backward coupling, the affiliation of neurons to assemblies can be re-defined by modifying the temporal synchronizations in the retinotopic layers. We demonstrate the effect of top-down influence on figure/ground separation and pattern recognition for different visual objects and scenes in computer simulations. It improves the segmentation of images containing multiple objects, including overlapping patterns of equal brightness, and it improves the rate of recognition. It is probable that our model can also be modified for other sensory modalities and it is proposed that selective feedback routing is one of the possible roles the numerous cortical backward projections may play.

Flexible Linking of Visual Features by Stimulus-Related Synchronizations of Model Neurons

Our models of visual information processing are based on the hypothesis that synchronized activities of sensory neurons serve to define perceptual relations: the features represented by the synchronized neurons are assumed to be linked and, thus, integrated into a perceptual entity. Recently, we found stimulus-related synchronizations in cat visual cortex that could play such role. These results are presented in chapter 2, together with discussions of the following questions: 1. What are the visual situations where stimulus-related activities in the visual cortex do become synchronized? 2. Where and by which neural mechanisms are synchronizations generated? 3. What possible roles do the synchronizations play in visual processing?

Invariances in Pattern Recognition

Humans and animals can recognize objects independent of object position in the visual field, largely independent of viewing angle and distance, and even independent of considerable variations in object shape. For object classification in the visual system, a comparison of sensory information with data in memory is required. Due to the large number of possible retinal pictures that can represent members of the same object class and even one individual object, such comparison is only feasible if an efficient object description can be generated that is object specific and invariant to changes in “non-essential” parameters. Of particular interest for models of invariance operations in the CNS are shift-invariant transforms. Shift-invariance is a primary invariance and other invariances can be derived from it. Size- and rotation-invariance, e.g., can be realized via a shift-invariance mechanism in combination with a logarithmic polar coordinate transform. With some approximations, the mapping of visual space to area 17 in the visual cortex of primates can be described by a logarithmic polar coordinate transform. Models of size-invariant processing in the CNS based on this mapping function have been proposed. In this paper, concepts for the generation of shift-, size-, and rotation-invariant pattern representations in the visual system are discussed, and a critical evaluation of their advantages, drawbacks, and neurophysiological implications will be given. Particularly, the paper will focus on two basic problems: (a) Neural networks are not well suited to perform exact arithmetic operations, as required in many models. A shift-invariant transform will be described that puts minimal demands on the neural transfer function. (b) Invariance operations generally require that the object has been separated from the background. Our model of region labeling via correlated neural activities preserves individual object contributions in composite pattern transforms, and is thus able to cope with the problem of figure/ground separation.