Beamforming (BF) architecture, an emerging approach for large‐scale antenna arrays, is a key task for the next decade's communication systems. The proposed approach in millimeter wave transmission is a fully connected analog phase‐shifter‐based BF architecture with limited radio frequency chains and imperfect channel state information (CSI). Deep learning (DL) is a powerful method for channel estimation and signal identification in wireless communications. Hence, this research proposes a DL‐enabled beamforming neural network (BFNN) which can be programmed to optimize the beamformer to attain better spectral efficiency. Simulation findings reveal that the proposed BFNN achieves significant performance gain and high robustness to imperfect CSI. The proposed BFNN greatly decreases the computational complexity by 0.16 million floating point operations (FLOPs) over 0.26 million FLOPs by conventional BF algorithms.