The difference between original data and its reconstructed version from an autoencoder, used to identify anomalies or unusual patterns.