It is shown that a neural network can be trained to predict the presence of
errors in turbo-decoded data, The inputs to the network are samples of the
cross-entropy of the component decoder outputs at two or more time instant
s. Such a neural network can be used as a trigger for retransmission reques
ts at either the beginning or at the conclusion of the decoding process, pr
oviding improved reliability performance and lower average decoding complex
ity than turbo decoding with CRC error detection.