Ontogenesis in Neural Networks

Lund University Cognitive Studies 24
ISSN 1101-8453
ISRN LUHFDA/HFKO--5018--SE 9 page

Abstract	When working with a neural network, it is advantageous to let it extract as much information as possible from the environment. We also want the network to perform better at this task. One way of achieving this is to make the perceptions of the network dependent on its current knowledge. This means that the network should be able to find structures in structures or to make categories of categories. This paper is an attempt to accomplish these goals by crafting the units to be feature correlators rather than feature detectors. A central component in the proposed construction is the use of spontaneous (or background) activity of the units as an important influence on the process.

Full paper	Full paper in pdf format.