Given the limited transmit power, the performance of over-the-air FEEL is often constrained by the device with the poorest channel quality. In this chapter, we utilize a RIS to relieve the communication bottleneck and provide a theoretical characterization of the convergence upper bound to quantify the negative impact of the cumulative aggregation error across all rounds. This motivates us to formulate a time-average transmission distortion minimization problem involving the joint optimization of the transceiver and RIS phase shifts, followed by presenting a graph neural network (GNN) based learning algorithm that directly maps channel coefficients to the optimized network parameters. Because of the permutation equivariance and invariance properties, the proposed algorithm has the advantage of low computational complexity and high algorithmic scalability. Results demonstrate that superiority of the proposed GNN approach for RIS-assisted FEEL.