In previous study, deep learning and autoencoder have been applied for data detection of NOMA systems, rather than the resource allocation of OFDMA/NOMA systems. In previous work, we proposed the use of non-deep-learning-based cross-layer resource allocation for OFDMA/NOMA video communication systems. In this paper, we apply a deep neural network and supervised learning to an OFDMA subcarrier assignment and NOMA user grouping problem in downlink video communication systems. The resource allocation results from our previous work are used as training data at the training stage. At the testing stage, we propose a conversion algorithm to map the result of the sigmoid activation function (values between [0,1]) of the output layer of the DNN to either zero (unassigned) or one (assigned), in order to meet two hard constraints. The PSNR performance is very close (within 0.2dB) to that but has lower complexity, due to the non-iterative approach used in the testing stage of the DNN.

Prior works either considered outage capacity of wireless video transmission systems but did not consider NOMA which is a key technology for 5G ultra-reliable low-latency (URLLC), or concerned the ergodic capacity of NOMA-OFDMA systems but did not consider the outage capacity emphasized in 5G URLLC scenario. In this paper, outage capacity (as well as ergodic capacity) maximization in a 5G URLLC scenario, are considered, using two proposed resource management schemes (i.e. B, C) and finally, proposed deep-learning-based versions of Schemes B and C (i.e. Schemes B' and C') to reduce the complexity and the latency for 5G URLLC. The proposed schemes re-allocate subcarrier according to outage capacityinstead of ergodic capacity only-maximization objective and choose the candidate user to gain subcarrier in a new way to improve the outage capacity. The numerical results show the proposed Scheme B and C increase the outage capacity (the percentage of satisfied users) from 79.2% for a prior work (Scheme A) to 85.2% and 92.6%, respectively. Scheme C also increases the ergodic capacity (average PSNR) from 34.7dB for Scheme A to 35.5dB. The deep learning based Schemes B' and C' perform slightly poorer than the corresponding non deep learning based Schemes B and C but the execution time/latency of Scheme B'/C' is less than Schemes B/C.