Visualizing Data using t-SNE
Laurens van der Maaten, Geoffrey Hinton2008JMLR 2008
Read it on jmlr.org(opens in a new tab)Why this one
Read this before you believe a cluster plot in a model eval, a vector database demo, or a founder deck. The useful idea is not that high dimensional data has a pretty 2D shadow. It is that visualization is an optimization problem with a bias: preserve local neighborhoods, exaggerate separation, and pay for that choice with distorted global geometry. People get t-SNE wrong when they treat distance between islands as semantic distance, or when they compare two plots without caring about perplexity, initialization, sample size, and random seed. Read this after Efficient Estimation of Word Representations in Vector Space if embeddings are your object of study, and before Visualizing Data using t-SNE becomes a tool you casually run. It will make you slower to claim insight from pictures, which is exactly the point.
What to take away
- t-SNE turns pairwise similarities into probability distributions, then fits a 2D map by minimizing KL divergence.
- Perplexity is not decoration: it sets the effective neighborhood size the plot tries hardest to preserve.
- Clusters and gaps can be artifacts of the objective, so inspect parameters and neighbors before naming structure.
Reads with
- Efficient Estimation of Word Representations in Vector Space
embeddings give t-sne the high-dimensional neighborhoods it is usually asked to explain
- Learning Transferable Visual Models From Natural Language Supervision
clip makes cross-modal embedding maps tempting, and t-sne shows how those maps can mislead
- Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
rag turns neighborhood geometry into product behavior, not just a visualization