This paper considers a two-hop multi-user multiple-input multiple-output (MU-MIMO) relay system, in which multiple users send information symbols to a multi-antenna base station (BS) with one-bit analog-to-digital converters via layered relays, each with one-bit transceiver. In this system, the maximum likelihood (ML) detection method is first proposed under the premise that perfect and global channel state information (CSI) is available at the BS. In practice, however, the CSI assumption is infeasible due to the high quantization noise effect. To overcome this limitation, a supervised learning-based data detection method is proposed by modeling the end-to-end system as effective parallel binary symmetry channels (BSCs). This detection method only exploits implicit CSI at the BS, which is simply learned by a limited amount of pilot transmissions. Lastly, a detection method using a deep neural network is presented, which is capable of removing the BSC model errors, while increasing the computational complexity for the parameter learning. Using simulation results, the detection performances of the three proposed detection methods are compared with some discussions.