Multiple Semantics.

[wiibo]

Would you please tell me about how to train the boundaries to manipulate multiple semantics?
is it putting two trained boundaries together or computing the covariance of diferent semantics' scores?

[songwoh]

As far as I understand, for each of the semantics, there is going to be different SVM linear classifier and the corresponding normal. Then as the paper suggests in section 2.2, you perform n1-(n1^T n2)n2 to obtain a new normal that you can add on to the original latent code to manipulate semantic n1 without affecting n2.
So in short, train 2 independent SVM classifiers with different semantic attribute scores, and the semantic scores is going to be a scalar, for each of them.