t-SNE

An exploration of t-distributed stochastic neighbor embedding (t-SNE) as generative mechanism. t-SNE is a widely used nonlinear dimension reduction technique. In the spring of 2021, while working on a lecture about this and similar methods, I noticed that t-SNE could generate point structures that looked quite intriguing. More recently, Lior Pachter has argued that t-SNE and the related UMAP do not serve a meaningful purpose in data analysis and are only useful for producing art.

The full t-SNE series is long-form generative art consisting of 99 pieces corresponding to consecutive random seeds 1 through 99. Here I display a curated subset of these 99 pieces. The entire series is available as NFTs on OpenSea.

All pieces in this series are characterized by seven distinct parameters. The meaning of these parameters is illustrated below.

2 groups

3 groups

4 groups

5 groups

Dark

Warm

Cold

Green

Purple

Brown

One

Multiple

Uniform

Noisy

Constant

Variable

1 panel

3 panels

4 panels

6 panels

8 panels

9 panels

None

Low

Medium

High