Bit-detection with neural networks from high-density magnetic recordings: a comparison

In high recording densities on magnetic disks, neighboring symbols overlap substantially, thus resulting in a hard-to-interpret read signal. Neural network architectures were developed by others to correctly detect the recorded bits. Even though the proposed neural network architectures contain feedback loops, testing was reported only for “single-bit errors”, meaning correct bits were used as feedback rather than the detected bits. The paper evaluates the proposed network architectures in a more realistic setting and addresses the issue of whether the performance gain due to specialization outweighs the performance loss due to sequence-errors (errors caused by the feedback of earlier errors)