t-SNE (t-distributed Stochastic Neighbor Embedding) Explained

t-SNE (t-distributed Stochastic Neighbor Embedding)

t-SNE is a nonlinear dimensionality reduction algorithm that converts high-dimensional data into 2D or 3D visualizations while preserving local similarity structure.

t-SNE is a dimensionality reduction technique used to visualize high-dimensional data in lower dimensions, typically:

• 2D
• 3D

It is widely used for:

• clustering visualization
• embedding visualization
• feature space analysis
• latent space exploration

Core Idea

t-SNE preserves:

• local structure
• neighborhood similarity

Points that are close in high-dimensional space remain close in lower-dimensional space.

t-SNE Visualization Example

flowchart LR

    A1[Cat Images]
    A2[Dog Images]
    A3[Car Images]

    A1 --> B[t-SNE Projection]
    A2 --> B
    A3 --> B

    B --> C[Clustered 2D Visualization]

Why t-SNE is Needed

High-dimensional data is difficult to visualize directly.

Examples:

• Word embeddings
• Image embeddings
• Transformer hidden states
• Feature vectors

t-SNE converts:

$$\text{High-Dimensional Space} \rightarrow \text{Low-Dimensional Visualization}$$

t-SNE uses heavy-tailed distribution to:

• avoid crowding problem
• separate distant clusters better

Applications of t-SNE

• NLP Embeddings
• Image Feature Visualization
• Transformer Embeddings
• Clustering Analysis
• Latent Space Visualization
• Anomaly Detection

High-Level Workflow

flowchart TD

    A[High-Dimensional Data] --> B[Compute Pairwise Similarities]

    B --> C[Convert to Probability Distribution]

    C --> D[t-SNE Optimization]

    D --> E[2D or 3D Embedding]

    E --> F[Visualization]

Example

Suppose each image is represented by:

$$x_i \in \mathbb{R}^{512}$$

t-SNE reduces it into:

$$y_i \in \mathbb{R}^{2}$$

for visualization.

---

Step 1: Similarity in High-Dimensional Space

t-SNE computes probability similarity between points.

Probability that point $x_j$ is neighbor of $x_i$:

$$p_{j|i} = \frac{ \exp(-||x_i - x_j||^2 / 2\sigma_i^2) }{ \sum_{k \ne i} \exp(-||x_i - x_k||^2 / 2\sigma_i^2) }$$

Where:

• $x_i$ = data point
• $\sigma_i$ = variance parameter

Step 2: Similarity in Low-Dimensional Space

Low-dimensional similarity uses Student t-distribution.

$$q_{ij} = \frac{ (1 + ||y_i - y_j||^2)^{-1} }{ \sum_{k \ne l} (1 + ||y_k - y_l||^2)^{-1} }$$

Where:

• $y_i$ = low-dimensional embedding

Optimization Objective

t-SNE minimizes divergence between:

• high-dimensional similarity
• low-dimensional similarity

Using KL Divergence:

$$KL(P || Q) = \sum_{i \ne j} p_{ij} \log \frac{p_{ij}}{q_{ij}}$$

Optimization Flow

flowchart TD

    A[High-Dimensional Similarities P] --> C[KL Divergence Loss]

    B[Low-Dimensional Similarities Q] --> C

    C --> D[Gradient Descent]

    D --> E[Updated Embeddings]

Important Hyperparameters

Parameter	Purpose
Perplexity	Controls neighborhood size
Learning Rate	Optimization step size
Iterations	Number of optimization steps
Dimensions	Output dimension (2D/3D)

---

Perplexity

Perplexity balances:

• local structure
• global structure

Typical values:

$$5 \le \text{Perplexity} \le 50$$

Advantages

• Excellent visualization quality
• Preserves local neighborhoods
• Works well with embeddings
• Reveals hidden clusters

Limitations

Limitation	Description
Computationally expensive	Slow on large datasets
Not deterministic	Different runs vary
Poor global distance preservation	Far clusters may distort
Primarily visualization tool	Not ideal for downstream ML

---

t-SNE vs PCA

PCA	t-SNE
Linear reduction	Nonlinear reduction
Fast	Slower
Preserves variance	Preserves neighborhoods
Good for preprocessing	Good for visualization

PCA + t-SNE Pipeline

Common workflow:

flowchart TD

    A[High-Dimensional Data]

    A --> B[PCA Reduction]

    B --> C[t-SNE]

    C --> D[2D Visualization]

PCA first reduces noise and dimensionality before applying t-SNE.

t-SNE vs UMAP

t-SNE	UMAP
Better local structure	Better global structure
Slower	Faster
More computationally expensive	More scalable
Widely used historically	Increasingly popular

---

---

Related Posts

• Dimensionality Reduction in Machine Learning — why high-dimensional data needs to be reduced before visualization
• Principal Component Analysis (PCA) — PCA for linear compression; t-SNE for non-linear visualization — use PCA first to reduce to ~50D, then t-SNE to 2D
• K-Means Clustering — moving from dimensionality reduction to clustering