Cutting off the outer layers of a neural network and appending fresh layers in order to train new inputs confuses me. This is how I understand transfer learning.
I want to know if the new input node corresponds to that of the originally pretrained network or if the new input node will start from the middle (where the old network was truncated).
From previously learnt theory, training the newly appended layers will also update the weights of the previously trained model. I suspect that this could work if there was a way to stop the original layers from updating while updating only the newly appended layers.
If the above logic is correct, after the transfer learning is complete, asking the model to predict objects from the current data will work. but asking the model to predict from a dataset which was previously used to train the model (before transfer learning happened), such input will have to pass through the new layers to arrive at an output. And I am afraid these new layers will not act well on the old data because they were not trained on the old data.
How does transfer learning circumvent such constraints?
Please, pardon me if I am not clear enough. Deep learning is really difficult to explain in writing. Pictures and words do a better job.