Softening learning procedure for the layered feed-forward networks with multiple output nodes

When we encounter a new case (a new input/output relationship), we will first check if the knowledge we obtained so far could interpret it. If yes, there is no further learning effort involved. If no, we might cram this unfamiliar case; then meditate to reason out a way of integrating it into our knowledge. Here I present a learning algorithm, referred as the softening learning algorithm, that imitates this way of learning in human beings.