2
$\begingroup$

I'm trying to solve a problem in the field of transfer learning, more specifically, domain adaption where both the source domain and target domain are labeled. Basically it's to predict the orientation of different material, the output is 2 angles that indicate the orientation of fibers.

I know that deep metric learning is a powerful method to solve the problem when data is sparse. But I checked some classic algorithms such as Triplet and center loss. They mostly concentrate on classification problems where an anchor and a center is required for each class. Is it possible to implement Metric Learning for regression problems so that I can transfer the knowledge learned by one material to the other?

$\endgroup$
1
  • $\begingroup$ I am really curious, did you find anything? I am looking for the same information. $\endgroup$
    – gipouf
    Dec 18 '20 at 18:45
1
$\begingroup$

You could have a look at the MLKR algorithm (Metric Learning for Kernel Regression): although the basic version learns a Mahalanobis metric (linear transformation), it should be easily adaptable to be able to use a DNN-based transformation instead (by using the same cost function which is the MSE of a soft-KNN regressor, but just using a DNN instead of a linear transformation to embed points in the transformed space in which the soft knn is computed)

$\endgroup$

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.