Design of adaptive and incremental feed-forward neural networks

The concepts of minimizing weight sensitivity cost and training square-error are applied on a biased two-layered perceptron using gradient descent to obtain an adaptive learning mechanism. Experiments show that the adaptive learning mechanism can tolerate noisy and inconsistent training instances by localizing the responses of conflicting data. Methods of resampling and dynamic normalization are introduced to construct an incremental feedforward network (IFFN) based on adaptive learning. This incremental learning mechanism has a measurable generalization capability and satisfies almost all of the six criteria proposed for incremental learning