Supervised training of neural networks via ellipsoid algorithms

It is shown that two ellipsoid algorithms can be used to train single-layer neural networks with general staircase nonlinearities. The ellipsoid algorithms have several advantages over other conventional training approaches, including explicit convergence results and automatic determination of linear separability, the elimination of difficulties associated with picking initial values for the weights, guarantees that the trained weights are in some acceptable region, certain robustness characteristics, and a training approach for neural networks with a wider variety of activation functions. Extensions to multilayer networks also exist.<
>