Dietmar Ruwisch

Parallel hardware implementation of Kohonen's algorithm with an active medium

We present a hardware realization of Kohonens algorithm using many simple, linearly coupled active elements, namely a reaction-diffusion medium, that is a spatially extended active element. This medium also controls the learning-process. Both classification and learning are realized in a way that only next-neighbor coupling is necessary. We exemplify the learning process by means of some simple tasks.

COLLECTIVE MODES OF THREE COUPLED RELAXATION OSCILLATORS: THE INFLUENCE OF DETUNING

The dynamics of a ring of three relaxation oscillators symmetrically coupled in the slow variable is considered both experimentally and by means of numerical simulations. It is shown that apart from the synchronous oscillation the basic set of periodic attractors typical for identical oscillators is not essentially affected by a small detuning. However, detuning creates a large set of new collective modes which are characterized by larger system periods and a variety of phase relations between the oscillators. Comparisons of the experimental data with different oscillator models reveal that the results do not crucially depend on the specific features of the models and possible experimental uncertainties.

New Digital Hardware Concept for Self-Organizing Feature Maps

We present a new, all digital hardware concept for Kohonen self-organizing feature maps. The neurons of this implementation are rather simple computation nodes. Aside from a data bus they are interconnected with next neighbors only. This low connectivity opens up the possibility of integrating a large number of neurons in a digital neuro-chip. Furthermore, a number of those neuro-chips could easily be linked to a neural network of arbitrary size.

Real‐time noise canceling based on spectral minimum detection and diffusive gain factors

In this paper, we propose a very simple but highly effective psychoacoustically motivated real‐time approach on the basis of spectral minimum detection and diffusive gain factors without a speech activity detector. The first processing step is the calculation of the short‐time power spectrum of the noisy speech signal. Estimating the background noise, the system calculates diffusive gain values in real time being obtained in a two‐layer structure: Each node of a layer is responsible for a single mode of the power spectrum. The first layer, called the ‘‘minimum detection layer,’’ holds the present noise level derived from the minimum of the input power spectrum which is detected within frames smaller than the FFT window. The minimum is transformed into a gain factor function using a signal‐to‐noise ratio control parameter. The diffusive gain factor interaction of neighboring modes is performed in the second layer, called the ‘‘diffusion layer,’’ in order to avoid ‘‘musical tones.’’ In the frequency domain,...