AN INTERACTIVE TEXTBOOK / OVERVIEW
The ideas behind the arithmetic.
A reference for channels, receptive fields, weight sharing, and what a prediction can—and cannot—tell you.
Loading the model and MNIST examples. Every calculation runs in your browser.
01 The input
28 × 28x ∈ [0, 1]
Draw a large, centered digit. Prediction updates as you draw.
Use an image file
02 Inside the network
LIVE TENSORSFollow the signal from pixels to a prediction. Select a layer to look inside.
Feature extraction ClassificationCH × H × W
Small patterns become bigger ideas.The first block detects local strokes. The second combines them across eight channels.
03 The prediction
10 classes–
The network thinks—softmax probability
A high probability is a model score, not a guarantee.
THE FIELD GUIDE
The ideas behind the numbers.
A little context for the things you just watched happen.
BUILT FROM THE ORIGINAL
Your PyTorch model, made explorable.
This lab follows Devin Grosko’s Forward architecture and the inspection ideas in Probe.py. The browser calculates the full forward pass, exposing the intermediate tensors instead of hiding them.