← AI Terminology
t-SNE
t-SNE is a nonlinear dimensionality-reduction technique that embeds high-dimensional points into 2D/3D by preserving local similarities, widely used for visualisation.
Famous for revealing clusters in embedding spaces.
Famous for revealing clusters in embedding spaces.
Why It Matters in AI
Understanding representations needs eyes. t-SNE maps make embedding geometry interpretable for NLP, vision, and biology — with caveats about global distances and parameter sensitivity.
Key Points
| Aspect | Description |
|---|---|
| Use | Visualise embeddings and hidden features |
| Idea | Match high-D neighbourhood probabilities in low-D |
| Limits | Not for quantitative distances; stochastic |
| Params | Perplexity critically affects layout |
| Related | UMAP, PCA |
| Practice | PCA to 50-D then t-SNE is common |
Simple Analogy
Redrawing a crowded high-dimensional party as a 2D floor plan that keeps friend groups standing near each other — great for seeing cliques, not exact road distances.
Common Usage Examples
sklearn.manifold.TSNE- Plot sentence embeddings by topic
- Colour by label to inspect separation
- Prefer UMAP for larger datasets often
Summary
In short: t-SNE visualises high-dimensional data in 2D/3D by preserving local neighbourhoods — the classic embedding inspection tool.