Closed dhx000 closed 3 years ago
Hi, sorry for the confusion . In fact, it is a typo. In camera-ready version, we have corrected it. I will also update the arxiv version soon.
In other words, there exists a dominant eigenvalue capturing the category information of the matrix . Therefore, given a sample mini-batch denoted by $\mathcal{X} = {x_i^k}$, we can obtain the corresponding latent features as $\mathcal{Z}$ through a posterior $q(z|x)$ parameterized by an encoder. By further conducting mode-1 flattening $\mathcal{Z}$ as $\mathbf{Z}$\footnote{We assume that the first dimension is associated with sample index.}, our proposed rank regularization can be given as $rank(\mathbf{Z}) = C$,
where $C$ is the number of categories of a specific task.
Thank you very much.
HI, the version in arxiv has updated
I have a question about your paper,The last third line on Page 3:‘’in other words,the rank of matirx is expected to be 1"?