This paper applies a recently developed neural network called plausible neural network (PNN) to function approximation. Instead of using error correction, PNN estimates the mutual information of neurons between input layer and hidden layer. The simple theory and training algorithm of PNN lead to a faster converging rate over that of feedforward neural networks. Experiment results confirm PNN has much better training performance. In addition, the bi-directional network structure of PNN provides the flexibility of approximating any attribute of the data within a single framework. As a result, PNN can compute a function and its inverse in the same network even the inverse function generally is a one-to-many mapping.