A Survey on Deep-Learning based Techniques for Modeling and Estimation of MassiveMIMO Channels

\textit{Why does the literature consider the channel-state-information (CSI) as a 2/3-D image? What are the pros-and-cons of this consideration for accuracy-complexity trade-off?} Next generations of wireless communications require innumerable disciplines according to which a low-latency, low-traffic, high-throughput, high spectral-efficiency and low energy-consumption are guaranteed. Towards this end, the principle of massive multi-input multi-output (MaMIMO) is emerging which is conveniently deployed for millimeter wave (mmWave) bands. However, practical and realistic MaMIMO transceivers suffer from a huge range of challenging bottlenecks in design the majority of which belong to the issue of channel-estimation. Channel modeling and prediction in MaMIMO particularly suffer from computational complexity due to a high number of antenna sets and supported users. This complexity lies dominantly upon the feedback-overhead which even degrades the pilot-data trade-off in the uplink (UL)/downlink (DL) design. This comprehensive survey studies the novel deep-learning (DLg) driven techniques recently proposed in the literature which tackle the challenges discussed-above - which is for the first time. In addition, we consequently propose 7 open trends e.g. in the context of the lack of Q-learning in MaMIMO detection - for which we talk about a possible solution to the saddle-point in the 2-D pilot-data axis for a \textit{Stackelberg game} based scenario.