I'm working on a computer vision classification subject using CNN. I use a pretrained model as basis of my model. I read in Stanford CS 231n class that for large dataset with domain specific, best practice was to fine-tune (retrain all layers).

I would like to know if fine-tune (with a right learning rate) always perform better than transfer learning or if there it is possible to erase some usefull features and overfit the dataset/domain.

  • $\begingroup$ It sounds to me that you're talking about two versions of transfer learning. In one case, you keep the same architecture and freeze the weights at some depth -- could be at "layer 0" so that you're adjusting all weights or could be deeper so you're not adjusting earlier weights. In the other case, you're chopping out entire layers and making new ones -- which might boil down to randomizing the weights of the transferred net. Is this correct? $\endgroup$ – Wayne Apr 11 '18 at 13:18
  • $\begingroup$ For me Transfer Learning is freeze features part and retrain softmax part and Fine-tuning is retrain whole network. $\endgroup$ – alexandre_d Apr 11 '18 at 14:07

Your Answer

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

Browse other questions tagged or ask your own question.