Discusses the representation of concepts and antecedent-consequent productions, and describes a method for inducing knowledge by abstracting such representations from a sequence of training examples. The proposed learning method, interference matching, induces abstractions by finding relational properties common to two or more exemplars. Three tasks solved by a program that performs an interference-matching algorithm are presented. Several problems concerning the relational representation of examples and the induction of knowledge by interference matching are also discussed. The similarities between this task and other computer science problems are indicated, and directions for future research are considered. 64 pp. Ref.