A class of self-generating neural networks




Edwards, Roderick
Illner, Reinhard
Leeming, Graeme

Journal Title

Journal ISSN

Volume Title



We discuss the design and behavior of families of neural networks which grow out of a single "mother" neuron in response to external stimuli and to the activities present in various parts of the net at a given time. The growth process is subject to a few fundamental rules, like - the ability of neurons to grow new neurons or connections is gradually exhausted with the number of generations - neurons are either of excitatory or inhibitory type - inhibitive neurons have a tendency to form long-range connections, whereas excitatory neurons "prefer" short-range connections. In addition, there are a number of free parameters in the equations driving the time evolution of the neural activities as well as the changes in the connection strengths. Our design is implemented using Matlab, such that the growth process of the network and its activity can be observed and controlled interactively on the computer screen.