t-SNE Flashcards
1
Q
What is t-SNE?
links:
https: //lvdmaaten.github.io/publications/papers/JMLR_2014.pdf
https: //www.youtube.com/watch?v=RJVL80Gg3lA
A
embedding technique commonly used for visualizing high-dim data in scatter plots
2
Q
similarity in stochastic neighbor embedding (SNE)
A
in SNE one computes a NxN similarity matrix in both the original and the low-dim data spaces; each matrix forms a probability distribution over pair of objects:
- similar objects’ pairs have high probability
- dissimilar objects’ pairs have low probability
3
Q
probability distributions in SNE
A
represented usually by normalized Gaussian or Student-t kernel
4
Q
learning the SNE
A
low-dim embedding is learned by minimizing the KL between the 2 distributions wrt to the locations of the points in the embedding