t-SNE

Dimensionality reduction: Visualize 64-dimensional handwritten digits in 2D/3D

Speed:iter/step

Dataset: MNIST Digits

📊 300 handwritten digit samples

📐 64 dimensions (8×8 pixels)

🏷️ 10 categories (digits 0-9)

Each point represents an 8×8 grayscale image of a handwritten digit. t-SNE maps these 64-dimensional vectors to 2D/3D while preserving local neighborhoods.

Perplexity: 30

Balances local vs global structure. Low values (5-15) emphasize local clusters; high values (30-50) reveal global relationships. Typical: 30.

Learning Rate: 200

Gradient descent step size. Too low: slow convergence. Too high: unstable. Range: 10-1000. Typical: 100-500.

Algorithm Phases

Early Exag.First 250 iterations amplify distances to separate global clusters

OptimizationRemaining iterations refine local structure while maintaining global layout

CompleteSimilar digits cluster together; dissimilar digits separate

What to Look For

• Digit clustering: Similar digits (e.g., 3s, 8s) naturally group together

• Confusion zones: Look for overlaps between similar digits like 4/9 or 3/8

• Handwriting variation: Each digit cluster shows natural writing style diversity

💡 Tip: Try different perplexity values to see how it affects local vs global structure visibility!