Computationally-improved optimal filtering for supervised learning [feedforward neural nets]
S. Benromdhane, F.M.A. Salam · 2002
We propose a modification of the Kalman filtering approach to supervised learning that avoids existing approximations while improving its overall computational efficiency. The modification eliminates the necessity of computing an inverse. The same global network structure is retained while the computational effort is extensively reduced. The performance of the approach with the modification, assessed from several test cases, is found to be more refined than the existing approaches.>