Incremental Learning with Optimal Generalizing Ability in Neural Networks
Abstract of paper published in Technical Report of The Institution of Electronics, Information and Communication Engineers (IEICE) NC95-9(1995-05)
For a given set of training data, a method of learning for optimally generalizing neural networks using functional analytic approach already exists. Here, we consider the case when additional training data is made available at a later stage. We devise a method of carrying out optimal learning with respect to the entire set of training data (including the newly added one) using the results of the previously learned stage. This ensures that the learning operator and the learned function can both be computed incrementally, leading to a reduced computational cost. Finally, a result on sampling to further reduce computational cost has been derived.