t- SN (t- Distributed Stocunec Sousedka Embedding) is a popular technique for visializing high- dimensional data in two or three dimensions. It helps reveal patterns and clusters that are not easily observable in te original data. Understanding thee communal principles behind t- SNE can improve its application and interpretation.

Core Concepts of t-SNE

t- SNE converts high- dimensional data pointes into a probability distribution that reflects their similarities. It then seeks a low- dimensional embedding that reserves these simarities as closely as possible. Thee process endives two main steps: computing pairwise similarities and minimizing a divergence compeen distributions.

Matematikal Foundations

In that e high- dimensional space, thee similarity between ein two point is modeled using a Gaussian distribution. Thee probability that point appli1; physi1; PY3; PY3; PY1; PY1; PY3; PY3; PY3; PY3is a PALIBOR of point physi1; PY3; PY3; PY3; PY1; PY1; PY3; PY3; PY3; PY3is given by:

p CLAS1; CLAS1; CLAS1; CLAS3; db; dc; dd; dd; dd; dd; dd; dd; dd; dd; dd; dd; dd; dd; dd; dd; dd; dd; dd; dd; dd; dd; dd; dd; dd; dd; dd; dd; dd; dd; dd; dd; dd; dd; dd; dd; dd; dd; dd; dd; dd) dd; dd; dd) dd d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d) d) d) d d

This definites a probability distribution over souseds for each point. Thee joint probability p austral1; pfiedlo1; FLT: 0 pfiíklad 3; ij pfi1; pfiedload: 1 pfiíklad 3is symmetrized as:

p CLAS1; CLAS1; CLAS1; CLAS3; ij CLAS1; CLAS1; CLAS1; CLAS1; CLAS1; CLAS1; CLAS1; CLAS3; CLAS3; CLAS1; CLAS1; CLAS1; CLAS1; CLAS1; CLAS1; CLAS1; CLAS3; CLAS3; CLAS3; CLAS3; CLAS3; CLAS3; CLAS3; CLAS3; CLAS3; CLAS1; CLAS1; C1d

kde je 1; fl1; FLT: 0 pt. 3m; N pt. 1m; fl1m; flt: 1 pt. 3m; is thotal number of point. In thee low-dimensional space, simarities are modelede using a Student 's t- distribution with on e pt. of freedom:

q 'I1;' FLT: 0 '; FLT: 0'; FLT; ij 'I1; FLT: 1'; FLT 3; Frac (1 'I1;' I1; 'I1; FLT: 2' I3; 'II1; i' I1; FLT: 3 'I3;' II1; 'II1;' Y 'I1; FLT: 4' I3; 'II3;' III1; 'III1;' III1; 'IIII1;' I1; 'I1;' I1; 'I1;' I1; 'II1;' I1; 'I1;' III1; 'II1;' I1; 'II1;

Optimization Process

Te goal is to find low-dimensional point is authori1; FLT: 0 p3; y p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1 p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1; p1 p1); p1 p1) p1) p1) p1); p1 p1 p1 p1); p1 p1 p1) p1 p1 p1 p1) p1) p@@

KL (P 'I124; FL12 4; Q) = sum _ {i neq j} p' I1; FLT: 0 'I3; ij' I1; FL1; FLT: 1 'I3; II3; II3; III1; FLT: 2' II3; Ij 'I1; FLT: 3' I3; III3; IIF: 4 'I3; IJ' I1; IJ 'I1; IH: 5' IIII3;

This is achied courgh gradient descent, settingg thee positions of points in th the low-dimensional space to reduce divergence. Thee gradients are computed on the differences between een under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under under un@@

Conclusion

Understanding thee Azial basis of t- SNE involves grasping how similarities are modeled and how thee optimization aligns these similarities across dimensions. This foundation helps in tuning parametrs and interpreting visualizations effectively.