Abstract: Algorithm-based computers are programmed, i.e., there must be a set of rules which, a priori, characterize the calculation that is implemented by the computer. Neural computation, based on neural networks, solve problems by learning, they absorb experience, and modify their internal structure in order to accomplish a given task. In the learning process, the available information is usually divided into two categories, examples of function values or training data and prior information, e.g. smoothness constraint, or other particular properties . From the learning point of view, the approximation of a function is equivalent with the learning problem of a neural network. In this paper we want to show the capabilities of a neural network to approximate arbitrary continuous functions and to build a practical neural network to approximate a continuous function. We have made some experiments in order to confirm the theoretical results.