128-dim features from trained CNN
Training Progress: Epoch 0 / 10 Accuracy: -
Watch how embeddings organize as the network learns
Loading data...
Class: -
Highlighting class -
Samples - Visible -
Highlighted - Hovered -

Display Options

5
70

Filter Classes

What is MNIST?

MNIST is a dataset of 70,000 handwritten digits (0-9), each represented as a 28x28 grayscale image (784 pixels).

Two Views

Raw Pixels: Project the 784-dim pixel values directly.
CNN Embeddings: Project 128-dim features learned by a neural network.

Training Evolution

In CNN mode, use the timeline to watch how embeddings evolve during training. The network learns to separate classes as it trains!

Interactions

  • Hover - View the original image
  • Click - Highlight all points of that class
  • Filter - Toggle classes on/off
  • Play - Animate through training epochs

PCA

Principal Component Analysis finds orthogonal axes of maximum variance. Fast and deterministic, but linear - it cannot capture non-linear relationships.

t-SNE

t-Distributed Stochastic Neighbor Embedding preserves local neighborhoods using probability distributions. Non-linear and excellent for visualization, but computationally expensive.

UMAP

Uniform Manifold Approximation and Projection uses topological methods. Faster than t-SNE and preserves both local and global structure better.

Algorithm Type Speed Structure
PCA Linear Fast Global
t-SNE Non-linear Slow Local
UMAP Non-linear Medium Both

Click a row to switch to that algorithm.