MNIST CNN · Live Inference

Handwritten Digit Recognition

Draw a digit (0–9) or try a sample · Inspect the exact 28×28 input tensor

// Draw Input

Try:

Stroke 18px Auto-predict

Draw a digit (0–9) above

// Model Prediction

28×28 Input 28x28 CNN preview

Confidence —

Inference —

// Class Probabilities

Architecture

4-layer Conv2D CNN with BatchNorm, MaxPooling2D, and Dropout (0.25 / 0.5) trained on MNIST.

Preprocessing

Bounding-box crop, aspect-ratio 20×20 scale, center-of-mass 28×28 alignment, and Gaussian smoothing.

Scope & Limits

Rejects blank canvases, expects single centered digits (0–9), and exposes full 10-class softmax uncertainty.