This work develops a deep learning power control framework for energy efficiency maximization in wireless interference networks. Rather than relying on suboptimal power allocation policies, the training of the deep neural network is based on the globally optimal power allocation rule, leveraging a newly proposed branch-and-bound procedure with a complexity affordable for the offline generation of large training sets. In addition, no initial power vector is required as input of the proposed neural network architecture, which further reduces the overall complexity. As a benchmark, we also develop a first-order optimal power allocation algorithm. Numerical results show that the neural network solution is virtually optimal, outperforming the more complex first-order optimal method, while requiring an extremely small online complexity.