Not OP, but the assertion that the two datasets be similar is just too incomplete to be correct, considering the topic is rather nuanced.
As the article states, the majority of the hidden layers, particularly the early ones, are very general ones that usually work out to identify edges/colours/etc... that generalise very well.
Because you pop-off the last fully-connected layer, transfer learning a network that has been trained on something like imagenet still generalises quite well to something completely different like medical imaging of tissue samples, although distance between the two sets will still be your main limitation.
Before I get ahead of myself, I'll refer you to this paper that tests the degree to which transfer learning is generalisable [0]. They conclude that difference between the two datasets is the main limitation, but that those models still work out better than random features, which I think proves that it's sufficiently generalisable for most tasks.
I like the notion of 'context' for transfer learning.
Where context can be parameterized.
The idea is that you learn a general model from your available data and you are able to specialise that model to perform well when adapted to different contexts.
A simple example is when the different contexts are different cost matrices or different expected ratios of positives and negatives.
So instead of learning a classification model from your training data, you learn a ranking model. You can make and adapt the different classification models (thresholds on the ranking) depending on the context of where that model will be deployed.
So for example you learn a ranking model from pictures that ranks women above men. When you want a classifier that classifies pictures as men or women, you chose the threshold from your ranking model depending on the confusion matrices costs for the context of where the model is being deployed.
I think a cool research theme is to think of similar tools for other aspects of transfer learning.
4 comments
[ 4.8 ms ] story [ 21.7 ms ] thread> the new dataset should be similar to the orginal dataset and the new dataset should be much smaller.
I don't think this represents the capabilities of transfer learning very well.
As the article states, the majority of the hidden layers, particularly the early ones, are very general ones that usually work out to identify edges/colours/etc... that generalise very well.
Because you pop-off the last fully-connected layer, transfer learning a network that has been trained on something like imagenet still generalises quite well to something completely different like medical imaging of tissue samples, although distance between the two sets will still be your main limitation.
Before I get ahead of myself, I'll refer you to this paper that tests the degree to which transfer learning is generalisable [0]. They conclude that difference between the two datasets is the main limitation, but that those models still work out better than random features, which I think proves that it's sufficiently generalisable for most tasks.
[0] http://papers.nips.cc/paper/5347-how-transferable-are-featur...
The idea is that you learn a general model from your available data and you are able to specialise that model to perform well when adapted to different contexts.
A simple example is when the different contexts are different cost matrices or different expected ratios of positives and negatives.
So instead of learning a classification model from your training data, you learn a ranking model. You can make and adapt the different classification models (thresholds on the ranking) depending on the context of where that model will be deployed.
So for example you learn a ranking model from pictures that ranks women above men. When you want a classifier that classifies pictures as men or women, you chose the threshold from your ranking model depending on the confusion matrices costs for the context of where the model is being deployed.
I think a cool research theme is to think of similar tools for other aspects of transfer learning.