This paper proposed a deep learning based physical layer design in communication system by establishing convolutional neural network in autoencoder. The communication signals are transformed by the network using random numbers as inputs. The ideal outcome for the formation of signals are in the form of constellation map which is traditionally formed by standard 16QAM modulation. The deep neural network extracts the characteristics of communication signals by minimizing the loss and adjusting particular parameters. This simulation illustrates that under Gaussian white channel

the network could share the same accuracy as 16QAM modulation. Additionally

in many complex situations

it could be made more adjustable by fine tuning structure parameters.