T-SNE

t-Distributed Stochastic Neighbor Embedding (t-SNE) is a nonlinear dimensionality reduction technique developed by Laurens van der Maaten and Geoffrey Hinton in 2008. It is primarily used for visualizing high-dimensional data in two or three dimensions while preserving local structure — the relationships between nearby points in the original space.

The method belongs to the family of neighbor-embedding algorithms. Unlike linear methods such as principal component analysis, which preserve global structure (variance along orthogonal axes), t-SNE prioritizes local neighborhoods: if two points are close in the high-dimensional space, they should remain close in the low-dimensional embedding. This makes it especially effective for revealing clusters and manifold structure in data that linear projections would obscure.

The Algorithm

t-SNE operates in two stages. First, it computes pairwise similarities in the high-dimensional space using a Gaussian kernel. Each point is the center of a Gaussian distribution, and the similarity between two points is proportional to the probability that one would pick the other as a neighbor under that distribution. The perplexity parameter controls the effective number of neighbors each point considers — a trade-off between local and global structure.

Second, t-SNE initializes points randomly in the low-dimensional space and iteratively adjusts their positions to minimize the Kullback-Leibler divergence between the high-dimensional similarity distribution and a low-dimensional similarity distribution. The low-dimensional distribution uses a Student t-distribution with one degree of freedom (a Cauchy distribution), which has heavier tails than the Gaussian. The heavier tails solve a crowding problem: in high dimensions, the number of moderately distant neighbors grows exponentially, and a Gaussian kernel in the low-dimensional space cannot accommodate them without compressing the local structure. The t-distribution's long tails allow moderate distances to be represented without forcing local neighborhoods apart.

The result is an embedding in which clusters often correspond to genuine structure in the data, but the global arrangement of clusters may be meaningless. The distance between cluster A and cluster B in a t-SNE plot tells you almost nothing about their actual distance in the original space.

Limitations and Misuse

t-SNE is among the most misused tools in machine learning. Researchers routinely interpret t-SNE plots as if they were faithful projections of the full data geometry, when in fact the algorithm is explicitly designed to preserve only local structure. A cluster in t-SNE does not necessarily indicate a cluster in the data; it may be an artifact of the embedding. The shape and size of clusters are not meaningful. The distance between clusters is not meaningful.

More troublingly, t-SNE is non-deterministic and sensitive to hyperparameters. Different random seeds produce different embeddings. Different perplexity values reveal different structures. The algorithm can find clusters in uniform random data — a failure mode that has been demonstrated repeatedly but ignored routinely. When a researcher runs t-SNE on their data and sees a clean separation into clusters, they are often seeing what the algorithm was designed to produce, not what the data actually contains.

This connects to broader issues in reproducibility in machine learning and the epistemic dangers of visualization. A picture is worth a thousand words, but a misleading picture is worth a thousand false claims. t-SNE gives the appearance of insight without the guarantee of truth.

t-SNE is a powerful tool for exploration, but the machine learning community has transformed it into a tool for justification. The algorithm was designed to help researchers see structure; it is now used to claim that structure has been found. This is not a failure of the algorithm. It is a failure of the epistemic culture surrounding it — a culture that values visual appeal over methodological rigor and treats a pretty plot as evidence of a genuine discovery.