Making backprop visible

I had already built neural nets from scratch — regression, then classification — deriving every weight update by hand. But there was still a gap: I understood backprop on paper, yet I couldn't actually see it happening.

So I built the thing I wish existed while I was learning it.

Scroll, and the network builds itself

It's a scroll-driven, interactive walkthrough of the backpropagation algorithm. You scroll, and the network assembles in front of you — the forward pass lighting up layer by layer, the error appearing at the output, then the gradients flowing backward through the chain rule, one derivative at a time. Every ∂E/∂w, ∂E/∂x, ∂E/∂y is shown exactly where it lives in the network.

No framework. No React, no chart library doing the heavy lifting — just vanilla JavaScript drawing the whole network as SVG, KaTeX typesetting the real equations, animated signals traveling the edges, dark and light themes, fully offline.

The thing that finally clicked

Backprop isn't a mysterious algorithm. It's the chain rule, applied node by node, backward. Once I could watch the signal travel down the edges, every "but how do gradients even reach the early layers" question just... disappeared.

Same lesson as always: build it before you abstract it away.