by Tomoharu Iwata, Kazumi Saito, Naonori Ueda, Sean Stromsten, Thomas Griffths and Joshua B. Tenenbaum
This paper addresses the problem of visualizing posterior probabilities of topic models. We have scanned another paper on how to visualize the documents via a probabilistic graphical model. Actually this paper is the inspiration of the latter.
By introducing an objective function resembling SNE, this paper solves an easier problem: the p table in SNE is now given, instead of computed via binary searches; the q table has a much smaller size.
Actually the latter paper decouples the number of topics with the embeding dimensions directly (but coupled with the number of clusters). We may observe quite similar phenomenon in this model.
As SNE, this model is also optimized with a gradient-descent method (alternatively, though).
No comments:
Post a Comment