Adaptative time constants improve the dynamic features of recurrent neural networks

摘要

We consider in this paper the improved features of recurrent neural networks where we allociate to each neuron-like unit an adaptative time constant T(sub)i. In order to quantify the effects of the T(sub)i's on the network, we present new results using a latching test (to evaluate the long-term memory capabilities) and a study of the modification of the stability regions with the time constants. Finally, a practical application of the Mackey-Glass chaotic signal prediction is presented.