A DATA MODEL FOR THE MORPHOGENETIC NEURON

Abstract

Listen

Translate article

Research on neural and neurofuzzy architectures indicates the need for the introduction of a more general concept than that of the neural unit, or node, introduced in the pioneering work by MacCuUoch and W.H. Pitts, Bull. Math. Biophys. 5, 115-133, 1943. The neural unit that is widely used today in artificial neural networks can be considered as a nonlinear filter. From this basic unit so-called "integrated" neural architectures may be built, in which many different neural networks cooperate. In order to do research on such neural architectures a description language is needed in which an artificial neural network can be considered as a single, dedicated entity. On the basis of these considerations we propose in the present paper a generalization of the concept of neural unit, which will be denoted as morphogenetic neuron. The name "neuron" was adopted because the activation function of such a device is characterized, in the same way as in classical neural units, by a bias potential and by a weighted sum of suitable, in general nonlinear, functions of morphogenetic fields. The attribute "morphogenetic" was chosen because the data determine the weights of the elementary source fields which generate the morphogenetic field by the linear operation of superposition. With the morphogenetic neuron it becomes possible to automatically reproduce a network of conventional neural units implementing a given input-output transfer function, without the need for resorting to laborious methods of synthesis, such as supervised training. The concept of the morphogenetic neuron appears to be in line with recent research in neurobiology. For example the auditory place cells are believed to “compute” the azimuth position (left-right) and elevation position (up-down) of a sound source. The position is obtained by measuring the internal time difference between signals which are delayed with respect to each other because of the difference in path-lengths between the sound source and each ear. In this case two inputs control the same auditory place cells S. Nobuo, in: Dynamic Aspects of Neocortical Function, 1984. The two signals are fused to obtain values of the morphogenetic field functions for the azimuth and elevation positions. Possible deteriorations or distortions of the primary signals are not important, because the significance of the data is solely connected with the delay of the signals and the morphogenetic fields associated with the ITD. Classically, data fusion is motivated by the presumption that