ncku-talk2024-03-07

Colloquium

DATE 2024-03-07　16:10-17:00

PLACE 數學系館 1F3174教室

SPEAKER Dr. Seonghyeon Jeong（國家理論中心）

TITLE Boundedness and Existence of a Minimizer of t-SNE Algorithm with Mild Conditions on the Data Points

ABSTRACT t-SNE is a nonlinear dimensionality reduction algorithm which is a method to represent high dimensional data in low dimension such as $\mathbb{R}^2$ or $\mathbb{R}^3$. When we reduce the dimension of data, we lose some information, hence we have to choose which information we want to preserve. In t-SNE, we quantify similarity of data points using Gaussian and student t-distribution function, and try to find a point distribution in low dimension which preserves the similarity of data by minimizing KL-divergence $\sum p \log {p\over q}$. t-SNE shows clusters of data pretty well, and therefore it is used in many area. In this talk, I present some researches regarding the t-SNE algorithm and discuss boundedness and existence of a minimizer of the KL-divergence.

Colloquium

DATE	2024-03-07　16:10-17:00

PLACE	數學系館 1F3174教室

SPEAKER	Dr. Seonghyeon Jeong（國家理論中心）

TITLE	Boundedness and Existence of a Minimizer of t-SNE Algorithm with Mild Conditions on the Data Points

ABSTRACT	t-SNE is a nonlinear dimensionality reduction algorithm which is a method to represent high dimensional data in low dimension such as $\mathbb{R}^2$ or $\mathbb{R}^3$. When we reduce the dimension of data, we lose some information, hence we have to choose which information we want to preserve. In t-SNE, we quantify similarity of data points using Gaussian and student t-distribution function, and try to find a point distribution in low dimension which preserves the similarity of data by minimizing KL-divergence $\sum p \log {p\over q}$. t-SNE shows clusters of data pretty well, and therefore it is used in many area. In this talk, I present some researches regarding the t-SNE algorithm and discuss boundedness and existence of a minimizer of the KL-divergence.