The basic idea of traditional similarity-based learning is that a program takes a number of instances, compares them in terms of similarities and differences, and describes the concept as a set of attributes common to positive instances. However, a concept exists because it is necessary to discriminate the concept from other concepts. Therefore, the all attributes common to positive instances are not always important to describe the concept. The important attributes are the ones which are necessary to discriminate the concept from other concepts. In this paper, I propose a new learning method based on differences among concepts. This method extracts the important attributes by changing the weight which is given to each attribute. Moreover, the method how the acquired concepts are memorized is important, especially when a given object is recognized or a concept is associated with other concepts. Therefore, a network structure based on similarity is proposed as a knowledge representation method.