Storage systems have a strong need for substantially improving their error
correction capabilities, especially for long-term storage where the
accumulating errors can exceed the decoding threshold of error-correcting codes
(ECCs). In this work, a new scheme is presented that uses deep learning to
perform soft decoding for noisy files based on their natural redundancy. The
soft decoding result is then combined with ECCs for substantially better error
correction performance. The scheme is representation-oblivious: it requires no
prior knowledge on how data are represented (e.g., mapped from symbols to bits,
compressed, and combined with meta data) in different types of files, which
makes the solution more convenient to use for storage systems. Experimental
results confirm that the scheme can substantially improve the ability to
recover data for different types of files even when the bit error rates in the
files have significantly exceeded the decoding threshold of the ECC.