A biological approach to plasticity in artificial neural networks

Using some observations on biological network plasticity, an idea is introduced to develop a procedure for varying network topology by simply pruning off connections in an initial overdimensioned fully connected network. The topology of a neural network is dynamically adapted using only the information of the learning set. The method, although it simply eliminates connections from an initial fully connected without creating new synapses, presents characteristics which can resemble certain kinds of biological behavior, such as the so-called critical period.<
>