Integrator neurons for analog neural networks

It is shown that integrators with saturation can be used as neurons for analog neural networks. A nonincreasing potential function is defined for the network. Computer simulations show that the neural network works well in wider parameter regions. Therefore, it is possible to choose reasonable parameters, for example, to avoid influence of noise of a certain frequency range without degrading performance, if changes are allowed in processing time; this is not the case for neural networks constructed from amplifiers. The reason for the different performances of the two networks is discussed.<
>