NN-learning with backpropagation and adaptive filter techniques

Explains current attempts using Kalman filter techniques in order to develop fast learning algorithms for feedforward neural networks (NN). There are two main ingredients: Taylor-series linearization during the weight updating process and backpropagation of the output errors according to the backpropagation (BP-) algorithm. Pointing to several other filter-based algorithms, the author discusses in somewhat more detail the least-square-lattice (LSL) approach which uses QR-decomposition. The QR-LSL algorithm has several advantages which make it to a highly interesting candidate for an efficient NN-learning algorithm