qSNE - Quadratic rate t-SNE optimizer with automatic parameter tuning

2020-06-04, 2020-06-04
software
qSNE is a t-SNE implementation with potentially quadratic converge rate (vs. linear) and automatic perplexity tuning. The implementation solves the exact original t-SNE problem and not an approximate version. Major features: A quasi-Newton optimizer which allows quadratic vs. linear rate of converge and, consequently, gives a similar result in e.g. √1000 ~ 30 iterations versus 1000 A perplexity range instead of a fixed value can be specific, in which case the perplexity is automatically optimized Fully supports Intel SSE and AVX (optional) and POSIX threads for parallelization For the details about the method, please refer to our publication on the matter.