1
$\begingroup$

I have a segmentation model trained using 1,000 images that can predict 4 classes (dog, cat, mouse, elephant). I would now like to extend the model with a 5th class (horse).

Horses are present in the 1,000 images used for the first model, but not labelled. Ideally I don't want to go back through those 1000 images and additionally label horse. What is efficient way to extend the model and incrementally add the new classes without re-labelling all previously used images for the new class, or labelling new images for all 5 classes?

Improve this question
$\endgroup$
1
  • $\begingroup$ I understand that we are talking about a multilabel classification problem, is that correct? Also, do you have training data for horses? If so, what does it look like? $\endgroup$
    – noe
    Commented Feb 14 at 7:59

1 Answer 1

Reset to default
0
$\begingroup$

Let's say you now have new data with the 5 labels. In theory, there is nothing wrong with adding that data and retrain.

But, be careful about learning artifacts. If you have horses that are labelled and some that aren't, your model might get confused and learn some strange reasons why it should label some horses and not others.

I also don't believe relabeling old data would be such a problem. You could:

  1. Label new data with horses
  2. train a new model on only that data not the old one
  3. Apply that model to the old data, and see what it comes up with. You might be able to reuse predictions as labels.
  4. Rinse and repeat a few times until you have most of your data relabeled.
Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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