Drawing: Click and drag on the grid (Right-click to erase)

3D Controls: • Left drag = rotate • Right drag = pan • Scroll = zoom

Touch Controls
https://github.com/Ramakm/NN-visualizer

MNIST Neural Network Visualizer

Interactive Neural Network Visualization

This application visualizes a compact Multi-Layer Perceptron (MLP) trained on the MNIST handwritten digit dataset. Draw a digit and watch in real-time as activations propagate through the network layers.

How It Works:

  • Draw: Sketch a digit (0-9) on the 28×28 grid using your mouse
  • Observe: Watch how your drawing flows through the network layers in 3D space
  • Predict: See the probability distribution for each digit class

Default Network Architecture:

  • Input Layer: 28×28 pixel grid (784 neurons)
  • Hidden Layer 1: 64 neurons with ReLU activation
  • Hidden Layer 2: 32 neurons with ReLU activation
  • Output Layer: 10 neurons (digits 0-9) with Softmax

Color Coding:

  • Neurons: Blue tones indicate low/negative activations, warm coral colors show strong positive activations
  • Connections: Color intensity represents weight magnitude and contribution strength

Tips:

  • Try different drawing styles to see how the network responds
  • Use the sample digits for quick testing
  • Adjust advanced settings to customize the visualization
  • Explore the 3D view from different angles

Advanced Settings

24

Controls how many of the strongest incoming weights are displayed per neuron. Higher values may slow performance.

0.0000

Hides connections with absolute weight below this threshold. 0 shows all connections.

Adjusts the visual thickness of connection lines in the 3D view.

1.4

Smaller values create finer lines; larger values fill the grid faster.

95%

Determines how much a brush stroke increases pixel brightness.