Non-linear neural networks with external noise
J. Leo van Hemmen, Kazimierz Rza̧żewski · Journal of Physics A Mathematical and General · 1987
The performance of neural network models with arbitrary non-linearity and Gaussian external noise superimposed on the synaptic efficacies is analysed. The memory function, though surprisingly robust, gradually fades out as the noise level is increased. In the low-noise limit the best performance is at zero temperature. There is a noise range, however, where optimal performance is obtained at a non-zero temperature.