Reproduce “On the continuum limit of t-SNE for data visualization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jeff Calder, Zhonggan Huang, Ryan Murray, Adam Pickarski
Researchers rigorously established the continuum limit of the t-SNE algorithm, deriving a continuous variational energy that illuminates its dimension-dependent behavior, including well-posedness in 1D and ill-posedness in higher dimensions (d > m). This foundational mathematical model provides theoretical explanations for observed phenomena like cluster formation and sensitivity to hyperparameters.
Key insights
- The continuum limit of the t-SNE energy features a non-convex, logarithmically growing attraction term and a repulsion term whose specific mathematical form depends on the embedding dimension *m*.
- For *m* = 1, the continuum energy is well-posed and admits a unique Lipschitz continuous, strictly monotone minimizer, even with its non-convex attraction term.
- In the practical scenario of strict dimension reduction (*d* > *m*), the derived continuum energy is ill-posed and unbounded below, implying that the nonlocal energy for fixed *h* acts as a crucial regularizer.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.