A General Neural-Network Decoder for Error-Correcting Codes

When communication channels have noise, errors often occur in messages transmitted from the sender to the receiver. Error correcting codes add redundancy to the original messages at the sender end and help to recover the original messages at the receiver end. The state-of-the-art error-correcting codes rely on certain mathematical relationships between the message bits, and these mathematical relationships are used to decode the original messages if the messages are corrupted with noise. This paper suggests a general neural-network-based decoder that can translate a received message to its closest codeword. The proposed method does not rely on the mathematical relationship between the bits in the encoded message.