Transfer learning is a method of reusing an already trained model for another task . The original training step is called pre-training. The general idea is that, pre-training “teaches” the model more general features, while the latter final training stage “teaches” it features specific to our own (limited) data.