Effect of External and Internal Noise on the Performance of Chaotic Neural Networks.

摘要

Biological evidence suggests that information encoding in the form of oscillatory patterns is advantageous compared to convergent fixed-point type memories. Freeman's KIII model is an example of operational chaotic memory neural networks. Noise plays a constructive role in the model by maintaining and stabilizing aperiodic orbits. Gaussian noise components are injected to the model different locations: at the input channels and also at a centrally located internal node. Depending on the noise intensity and bias, resonance effects have been identified in the KIII model. The observed noise effects have some similarity with stochastic resonance but there are very essential differences The interaction of noise with the oscillatory signal has a resonance character in the KIII model. The oscillatory signal in KIII, however, is not coming from the external world, but it is the result of the interaction of various intern components. Therefore, the signal has an intimate interference with the noise. These effects are illustrated in pattern recognition problems.