2
$\begingroup$

Suppose a model M classifies apples and oranges. Can M be extended to classify a third class of objects, e.g., pears, such that the new images for 'retraining' only have pears annotated and apples and oranges ignored? That is, since M already classifies apples and oranges, can the old weights be somehow preserved and let the retraining focus specifically on learning about pears?

Methods such as fine-tuning and learning without forgetting seem to require all objects in the new images annotated though.

$\endgroup$
2
$\begingroup$

Yes this is standard transfer learning. Using a trained model, we can freeze the first N hidden layers of a classifier, except for the last few. This will allow our previous relevant training to be retained whilst also being able learn new features and target new classes.

We will then initialize a new output layer that works our new context(i.e sigmoid, 1 node for binary classifier). Everything is now set to resume training on our new y_targets.

Take a look at this link for some more info on transfer learning.

If you want no perturbance to your past learning, I would recommend freezing all previous hidden layers and then tacking some additional ones on.

$\endgroup$
1
  • $\begingroup$ But doesn't fine-tuning require even the old objects annotated in the new training images? I'm trying to avoid having to annotate the old objects in the new images. $\endgroup$
    – John M.
    Apr 7 '19 at 18:01

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.