Seeing high-dimensional data: t-SNE

Source lecture

StatQuest: t-SNE, Clearly Explained (Josh Starmer (StatQuest)). Watch on YouTube.

What you’ll learn

This is lesson 12 of Track 10, the close of Phase 3 (Finding structure without labels). By the end you will be able to read a t-SNE plot without over-reading it, trusting what it does say (which points cluster together) and resisting the urge to read meaning into what it does not (the distance between clusters, the size of a cluster, the exact arrangement on the page). The one capability to walk away with: looking at any t-SNE figure in modern AI, you can name which interpretations are valid and which are common misreadings.

The track structurally mirrors StatQuest’s intuition-first machine learning videos. Full attribution is in this lesson’s references.

Where this fits

This closes the unsupervised phase. PCA, the previous lesson, gave you a linear way to compress data, great at preserving variance but limited to straight axes. t-SNE is the nonlinear partner built specifically for visualization, picking up where PCA’s straight axes hide clusters that sit on curved surfaces. Together the two lessons cover the dimensionality-reduction half of unsupervised learning. From here the track turns to its last and most cross-cutting question: how do you know whether any model is actually any good? Phase 4 opens with the bias-variance tradeoff.

Before you start

Prerequisite: Lesson 11, Squeezing dimensions: PCA. You need the idea of dimensionality reduction and why visualizing high-dimensional data is hard, because this lesson is the nonlinear visualization counterpart to PCA. No math beyond comparing similarities; the t-SNE algorithm itself is presented at an intuition level.

By the end, you’ll be able to

• Explain what t-SNE is for (visualization) and not for (preprocessing)
• Describe at intuition level how it works (matching 2D similarities to high-D ones)
• Identify what t-SNE preserves (local) and does not (global)
• Avoid the three common misreadings of t-SNE plots
• Explain perplexity and the practice of running t-SNE multiple times

Time and difficulty

• Read time: about 12 minutes
• Practice time: about 15 minutes (a valid-vs-misreading exercise, a perplexity-tuning question, and flashcards)
• Difficulty: standard