direkt zum Inhalt springen

direkt zum Hauptnavigationsmenü

Sie sind hier

TU Berlin

Inhalt des Dokuments

Modelle neuronaler Systeme

Ausgewählte Abstracts

A model for the reaching reflex of an infant
Zitatschlüssel grue03
Autor Grünewälder, S. and Bibel, W. and Obermayer, K.
Buchtitel Proceedings of the 29th Göttingen Neurobiology Conference
Jahr 2003
Zusammenfassung In this presentation a model that imitates a behavior which occurs during a learning step of an infant will be examined. The model is based on a unsupervised learning algorithm that builds up on the hebbian learning rule and on the rescorla-wagner theory. The behavior during the learning step of an infant which the model will imitate is well analysed in the Developmental Psychology and known under the name reaching reflex. An infant learns within it's first sixth months of living how to reach successfully for an object. Therefore the learning task for the model is to construct a neuronal structure which activates muscles in relation to a visual input. The model operates in a simulated environment and it consists of a neuronal structure. The structure gets its main input through 40x20 visual neurons. In relation to the visual input the model has to activate motoric neurons to move a virtual arm consisting of two limbs. The neuronal net operates spike based. Therefore the model has to excite the motoric neurons with a specific spike pattern to reach successfully. The network consists of two different structures which create at different time the arm movement. In the beginning the movement is made through an unmodifiable neuronal feedforward networkstructure which creates spike patterns for the motoric neurons in dependency of environmental regions in which an object is. The resulting spike patterns which occur in the net are used as training examples for the learning algorithm to build up a new link structure which is the second structure that is used later from the model to create the arm movement. Spoken abstractly the learning algorithm extracts from the spike patterns causal relations between different spikes and successful reaching. It determines which spikes are relevant and which are not for success. The links are used to encode this information. They store a weighted spike pattern with temporal information. The weights are used to express the relevance of different spikes within a pattern. With this links it is possible to represent complex relations. It's for example possible to express that reaching is often successful if a visual neuron n1 and n2 creates spikes 10 ms before a motoric neuron m1 creates one. The task of the learning algorithm is to build up the links and to weight the encoded spikes with the associated timings. One main reason for using this new kind of link is that when a classical recurrent network structure would be used every causal relation needed to be expressed through an own subnetwork structure and the number of neurons in the net would explode. The movement of the arm with these links works in that way that the net tries to recreate the causality expressing spike patterns with the help of activation propagation through the links. The resulting movement is better than the movement through the initial structure.
Link zur Publikation Download Bibtex Eintrag

Zusatzinformationen / Extras


Schnellnavigation zur Seite über Nummerneingabe