What can memorization learning do from noisy training examples?

When we are concerned with a learning method, such as regularization learning, which does not directly deal with generalization error, we usually use it to achieve some "true objective learning". That is, we first have some objective such as minimization of generalization error, then we look for a learning method which could achieve the objective learning. There is, however, another situation. When we have developed a learning method, we wish to apply it to a wide range of different purposes. We discuss the latter problem. We clarify the bound of applicability of memorization learning within a family of projection learning. The bound is determined by the location of sample points and the nature of noise.