In this work, a new approach to detection in high density magnetic recording is presented. The basic detector structure is a bank of parallel sub-detectors which are specialized in different channel states. The individual adaptation of the sub-detectors to the channel properties is done by a learning algorithm so that no explicit channel model is required. On both simulated and experimental data it is shown that the proposed detection scheme has excellent performance in the presence of strong nonlinear distortion and signal dependent noise. At user densities of 3.5 and 4.0 using RLL(1,7) encoding the new scheme outperforms multi-level decision feedback equalization (MDFE) detection by a factor of up to 500 in error rate. The structure is welt suited for hardware implementation.