Training error is something we can always compute for a (supervised) learning algorithm. But what we want is the error on the future (unseen) data. We define the generalization error as the expected error of all possible data that could come in the future. We cannot compute it, but can approximate it with error computed over a testing set.