About Rank

[IvyYZ]

Thank you for your excellent work.
And then I have a question, can we replace the kernel norm by just taking the rank of the matrix, and then making its rank as close to the class as possible.

[cuishuhao]

I do not think so for 2 reasons.

1. The calculation of matrix rank is NP-hard problem. It is difficult to limit the rank.
2. Rank denotes prediction diversity as shown in the paper, simply restricting diversity will ignores prediction discriminability.

[IvyYZ]

Thank you for your reply