Alternative learning procedures for reedforward neural networks

The majority of currently used learning algortihms for feedforward networks is based on gradient descent methods of non-linear optimization. They have proven successfull in solving real-world problems, nevertheless they are known to suffer from several shortcomings (numerical instability, time demands, communication overhead). Several theoretical results concerning the approximation power of neural networks have been established in the last decade, but a little effort has benn made to use these results-based on functinal approximation theory-for proposal of novel learning procedures. We plan to carefully study the results and proof techniques of Kolmogorov, Sprecher, Kůrková, Leshno, Mhaskar and others to derive alternative learning procedures for feedforward networks. Numerical and computational properties of these algortihms will be studied by means of theory and experiments on standard benchmarks.