Most explanations stop at the diagram or jump straight to the calculus. I wanted the bit in between, so I animated it.
Six minutes, eight examples. One real input travels through a five-layer network, every weighted sum resolves into an actual number, the loss gets calculated, and the error flows back through every weight. Then the same machinery solves five other problems, which makes it clear how little actually changes between them.
No narration, everything is on screen, so pause wherever you like.
Every animation flags its own simplifications on screen. The neuron labels are illustrative – real networks learn their own messy features.
Leave a Reply