A Meta Learner Autoencoder for Channel State Information Feedback in Massive MIMO Systems

In massive multiple-input multiple-output frequency division duplexing systems, the user equipment should independently estimate the massive downlink channels state information and then feed them back to the base station. This process results in a large signaling overhead. Deep learning approaches tried to overcome this challenge using neural networks as an autoencoder to learn the mapping between the input and corresponding output. However, this type of learning consumes massive training datasets to learn. Also, it can not make use of the learning through the internal information within the tasks and thus, it can not reach the convergence quickly as its parameters are randomly initialized. In this paper, we introduce a meta learner-based autoencoder for tackling the feedback overhead. The proposed approach is mainly based on finding a good initialization of the parameters of the autoencoder to adapt rapidly to new tasks with a few number of samples. The results show that the proposed autoencoder based on the meta-learner method outperforms the state of the art with a margin.