Previous studies have demonstrated that end-to-end learning enables\nsignificant shaping gains over additive white Gaussian noise (AWGN) channels.\nHowever, its benefits have not yet been quantified over realistic wireless\nchannel models. This work aims to fill this gap by exploring the gains of\nend-to-end learning over a frequency- and time-selective fading channel using\northogonal frequency division multiplexing (OFDM). With imperfect channel\nknowledge at the receiver, the shaping gains observed on AWGN channels vanish.\nNonetheless, we identify two other sources of performance improvements. The\nfirst comes from a neural network (NN)-based receiver operating over a large\nnumber of subcarriers and OFDM symbols which allows to significantly reduce the\nnumber of orthogonal pilots without loss of bit error rate (BER). The second\ncomes from entirely eliminating orthognal pilots by jointly learning a neural\nreceiver together with either superimposed pilots (SIPs), linearly combined\nwith conventional quadrature amplitude modulation (QAM), or an optimized\nconstellation geometry. The learned geometry works for a wide range of\nsignal-to-noise ratios (SNRs), Doppler and delay spreads, has zero mean and\ndoes hence not contain any form of superimposed pilots. Both schemes achieve\nthe same BER as the pilot-based baseline with around 7% higher throughput.\nThus, we believe that a jointly learned transmitter and receiver are a very\ninteresting component for beyond-5G communication systems which could remove\nthe need and associated control overhead for demodulation reference signals\n(DMRSs).\n