# Stochastic Neighbor Embedding under f-divergences

arXiv: https://arxiv.org/abs/1811.01247 · HTML: https://mlsanity.com/paper/1811.01247/

## Key facts

- Cited by 9 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 3 Nov 2018.

## Built on by (most cited first, 9 of 9)

- [Manifold learning: what, how, and why](https://mlsanity.com/paper/2311.03757/) (2023)
- [Theoretical Foundations of t-SNE for Visualizing High-Dimensional Clustered Data](https://mlsanity.com/paper/2105.07536/) (2021)
- [Your Contrastive Learning Is Secretly Doing Stochastic Neighbor Embedding](https://mlsanity.com/paper/2205.14814/) (2022)
- [Convergence analysis of t-SNE as a gradient flow for point cloud on a manifold](https://mlsanity.com/paper/2401.17675/) (2024)
- [Random Forest Autoencoders for Guided Representation Learning](https://mlsanity.com/paper/2502.13257/) (2025)
- [t-SNE Exaggerates Clusters, Provably](https://mlsanity.com/paper/2510.07746/) (2025)
- [Dimensionality Reduction using Elastic Measures](https://mlsanity.com/paper/2209.04933/) (2022)
- [Towards a comprehensive visualization of structure in data](https://mlsanity.com/paper/2111.15506/) (2021)
- [Cluster and then Embed: A Modular Approach for Visualization](https://mlsanity.com/paper/2509.03373/) (2025)

## Abstract (arXiv)

The t-distributed Stochastic Neighbor Embedding (t-SNE) is a powerful and popular method for visualizing high-dimensional data. It minimizes the Kullback-Leibler (KL) divergence between the original and embedded data distributions. In this work, we propose extending this method to other f-divergences. We analytically and empirically evaluate the types of latent structure-manifold, cluster, and hierarchical-that are well-captured using both the original KL-divergence as well as the proposed f-divergence generalization, and find that different divergences perform better for different types of structure. A common concern with t-SNE criterion is that it is optimized using gradient descent, and can become stuck in poor local minima. We propose optimizing the f-divergence based loss criteria by minimizing a variational bound. This typically performs better than optimizing the primal form, and our experiments show that it can improve upon the embedding results obtained from the original t-SNE criterion as well.