rustograd · a neural net you can watch learn

This app trains a tiny neural network to tell two kinds of images apart, and draws it as a 3-D cube as it learns, step by step.

HOW TO USE

• Type a search term for class A and class B (e.g. “sunflower” vs “blue sky”).
• Hit Fetch & Train — example images are pulled from Wikimedia Commons and the net trains live.
• Drop any image into Test image to see which class it’s more like.
• Advanced options tune the network — hover any field for an explanation.

WATCHING IT LEARN

Each small cube is a neuron. ←/→ switch the color lens (gradient · weight · activation), ↑/↓ change the layout, drag to orbit, scroll to zoom, and hover a neuron for its numbers. On the gradient lens you’ll see the cube “cool down” as the loss falls.

WHAT TO EXPECT

It’s a toy. It does best with visually distinct classes (color) and will struggle with semantic ones like cat-vs-dog — it only sees a 16×12 thumbnail by default, with no convolutions or pretrained features. Small training sets mean it can overfit, too.

LEARN MORE

This is a Rust reimplementation of Andrej Karpathy’s micrograd. His YouTube video “The spelled-out intro to neural networks and backpropagation: building micrograd” is a great place to start learning how this works. :)

Start →