Digital receivers are required to recover the transmitted symbols from their\nobserved channel output. In multiuser multiple-input multiple-output (MIMO)\nsetups, where multiple symbols are simultaneously transmitted, accurate symbol\ndetection is challenging. A family of algorithms capable of reliably recovering\nmultiple symbols is based on interference cancellation. However, these methods\nassume that the channel is linear, a model which does not reflect many relevant\nchannels, as well as require accurate channel state information (CSI), which\nmay not be available. In this work we propose a multiuser MIMO receiver which\nlearns to jointly detect in a data-driven fashion, without assuming a specific\nchannel model or requiring CSI. In particular, we propose a data-driven\nimplementation of the iterative soft interference cancellation (SIC) algorithm\nwhich we refer to as DeepSIC. The resulting symbol detector is based on\nintegrating dedicated machine-learning (ML) methods into the iterative SIC\nalgorithm. DeepSIC learns to carry out joint detection from a limited set of\ntraining samples without requiring the channel to be linear and its parameters\nto be known. Our numerical evaluations demonstrate that for linear channels\nwith full CSI, DeepSIC approaches the performance of iterative SIC, which is\ncomparable to the optimal performance, and outperforms previously proposed\nML-based MIMO receivers. Furthermore, in the presence of CSI uncertainty,\nDeepSIC significantly outperforms model-based approaches. Finally, we show that\nDeepSIC accurately detects symbols in non-linear channels, where conventional\niterative SIC fails even when accurate CSI is available.\n