For decades, optimization has played a central role in addressing wireless resource management problems such as power control and beamformer design. However, these algorithms often require a considerable number of iterations for convergence, which poses challenges for real-time processing. In this work, we propose a new learning-based approach for wireless resource management. The key idea is to treat the input and output of a resource allocation algorithm as an unknown non-linear mapping and to use a deep neural network (DNN) to approximate it. If the non-linear mapping can be learned accurately and effectively by a DNN of moderate size, then such DNN can be used for resource allocation in almost real time, since passing the input through a DNN to get the output only requires a small number of simple operations. In this work, we first characterize a class of `learnable algorithms' and then design DNNs to approximate some algorithms of interest in wireless communications. We use extensive numerical simulations to demonstrate the superior ability of DNNs for approximating two considerably complex algorithms that are designed for power allocation in wireless transmit signal design, while giving orders of magnitude speedup in computational time.