Hi, is there any quantifiable scientific process to understand whether a deep learning model trained on a certain image classification dataset can be used for transfer learning for image classification on another dataset ?
Or rather how do we quantify that a certain image dataset is quite different than another dataset mostly for methods of classification, which makes it unsuitable for transfer learning ?
Hi Titas De , I am not aware of any standardized metric to estimate how well a pretrained network will do on a target dataset without trying.
Whereas to measure divergence between two datasets, I suggest using KL divergence normalized with self divergence as KL-D does not have a standard range. That should give you an idea about how close/far apart two image datasets are. And by trying this over multiple subsets (varying number of samples) of the target dataset, you can get an idea(quantitatively) of what your original network is capable of.
Hope this helps!
- Badges
- Science topic