MNIST Interactive Demo

About This Model

This model has been trained on the MNIST dataset with an accuracy of roughly 97.5% (250 prediction errors out of 10,000 test images).

Architecture: Multi-layer perceptron [ Dense [784 → 128], ReLU, Dense [128 → 10] ]

Weight initialization: He uniform distribution. Optimizer: Stochastic Gradient Descent (no momentum).

Trained entirely in-browser using WebGPU compute shaders.

Note: Your handwriting may differ from the MNIST training data, which could affect recognition accuracy.

Draw a Digit

Layer Activations

Visualizes data flowing through each layer as you draw. All layers are fully connected (784 → 128 → 128 → 10 neurons).

Learned Parameters

Weight and bias matrices from the pre-trained model, normalized to 0–1 range for visualization.