A method for noise filtering with feed-forward neural networks: Analysis and comparison with low-pass and optimal filtering

C. Klimasauskas (1989) discussed the use of layered, feedforward neural networks that use the back-propagation learning rule for noise filtering. An analysis is presented of a similar neural network method for noise filtering, and, using computer simulations, it is compared to predictions of the analysis and to the noise-filtering properties of Wiener (or optimal) filtering and low-pass filtering. The signals used for comparison are two chaotic signals and two random frequency mixtures of sine waves, all corrupted with additive white noise. The network method compares favorably when filtering chaotic signals, and its performance can be approximately predicted by a simple equation that is based on the network configuration. Even networks that are constructed poorly will do some filtering, although knowledge of the dimension of the attractor for the signal being filtered is useful for determining an optimal configuration