Most traditional theories presuppose abstract underlying representations ( URs) and a set of rules to explain phonological processes. There are, however, a number of questions regarding this approach: Where do URs come from? How are rules formed and related to each other? This paper proposes a new approach, "performance phonology", that does not require any explicit rules and URs. In this study, it is hypothesized that rules would emerge as the generalizations a connectionist network abstracts in the process of learning to associate forms with the meanings of the words and URs could emerge as a pattern on the hidden layer. Employing a simple recurrent network, a series of simulations on different types of phonological processes was run. The results show that this network is capable of learning various types of phonological phenomena without URs and explicit rules, thus providing the justification for "performance phonology".