Neural Network Explorer

The frustration that started this project was that every neural-net diagram I’d seen was either a marketing illustration or a paper figure. Neither lets you poke at it. I wanted something where you could feed an input, watch activations light up layer by layer, and develop the kind of intuition you only get from interaction.

Building it forced me to read TensorFlow’s introspection APIs more carefully than I ever had. Most of the surface area I’d ignored for years turns out to be useful, model graph traversal, intermediate tensor capture, that sort of thing. I came out of the project a better TF user even though the model itself was incidental.

The visual side was harder than I expected. Three.js makes the basics easy, but the moment you want a network with thousands of neurons rendered cleanly, you hit instancing, level-of-detail rendering, and a long fight with depth-buffer artifacts. I spent more time on “the network looks calm when you zoom out” than on the actual ML.

The bit I think landed best is the step-through animation. Pressing forward replays a forward pass one layer at a time, with the activations fading into view. It’s the closest I’ve gotten to making the math feel like motion. People who said they’d been confused by this stuff for years told me it clicked the moment they watched a single example pass through.

The part I’m least proud of is the educational copy. I think clear explanations of what a network is doing are genuinely hard to write, and mine are okay at best. If anyone wants to fork this and bring better pedagogy to it, please do. The hardest part of a teaching tool isn’t the visual. It’s the words around it.